The Dynamics of Coherent Vortices Near The Turbulent/non-Turbulent Interface Analysed by Direct Numerical Simulations

The dynamics of coherent vortices near the turbulent/non-turbulent interface
analysed by direct numerical simulations

Ricardo Jos Nunes dos Reis
Supervisor: Doctor Jos Carlos Fernandes Pereira
Co-Supervisor: Doctor Carlos Frederico Neves Bettencourt da Silva
Thesis approved in public session to obtain the PhD degree in Mechanical Engineering
Jury nal classication: Pass with Merit
Jury
Chairperson: Chairman of the IST Scientic Board
Members of the committee:
Doctor Jos Carlos Fernandes Pereira
Doctor Lus Miguel de Oliveira e Silva
Doctor Fernando Manuel Coutinho Tavares de Pinho
Doctor Joo Manuel Melo de Sousa
Doctor Carlos Frederico Neves Bettencourt da Silva
Doctor Miguel ngelo Cortez Teixeira
2011
The dynamics of coherent vortices near the
turbulent/non-turbulent interface
analysed by direct numerical simulations
Ricardo Jos Nunes dos Reis
Supervisor: Doctor Jos Carlos Fernandes Pereira
Co-Supervisor: Doctor Carlos Frederico Neves Bettencourt da Silva
Thesis approved in public session to obtain the PhD degree in Mechanical Engineering
Jury nal classication: Pass with Merit
Jury
Chairperson: Chairman of the IST Scientic Board
Members of the committee:
Doctor Jos Carlos Fernandes Pereira, Professor Catedrtico do Instituto Superior Tcnico,
Universidade Tcnica de Lisboa
Doctor Lus Miguel de Oliveira e Silva, Professor Catedrtico do Instituto Superior Tcnico,
Doctor Fernando Manuel Coutinho Tavares de Pinho, Professor Associado (com Agregao) da
Faculdade de Engenharia, da Universidade do Porto
Doctor Joo Manuel Melo de Sousa, Professor Associado do Instituto Superior Tcnico,
Doctor Carlos Frederico Neves Bettencourt da Silva, Professor Auxiliar do Instituto Superior
Tcnico, Universidade Tcnica de Lisboa
Doctor Miguel ngelo Cortez Teixeira, Investigador Auxiliar da Universidade de Lisboa
Funding Institutions
Fundao para a Cincia e a Tecnologia
2011
Abstract
Free shear ows are present in a large number of natural and engineering ows. One of their distin-
guishing features is the presence of a sharp interface separating the turbulent (T) and the non-turbulent
(NT) ow regions. The physical processes happening in this interface play a pivotal role in determining
the spreading and/or mixing rates in shear layers (turbulent entrainment (TE)).
The study of the T/NT interface has got a new impulse in the last ten years, with experimental and
numerical results challenging many previous assumptions, e.g., , showing that TE is mainly a viscous
process acting at the small scales (nibbling), contrary to the classical large scale engulfment theory.
This thesis covers new ground on this subject, studying the role of coherent vorticity structures near the
T/NT interface.
The characteristic sizes and dynamics of the coherent structures were studied near the turbulent/non-
turbulent (T/NT) interface. The results obtained shed a new light on the nibbling eddy vortices responsi-
ble for the TE in jets.
Additionally, to further the study presented, a spatial Direct Numerical Simulation (DNS) code for
plane turbulent jets was ported to a parallel architecture and two massive simulations achieved and
reported.
Keywords: Turbulent Jets, Turbulent Entrainment, Coherent Structures, Turbulent/Non-turbulent
Interface, Direct Numerical Simulation, High Performance Computing, Parallel Computing
Resumo
Os escoamentos de corte livre esto presentes num grande nmero de contextos naturais e de
engenharia. Uma das suas principais caractersticas a presena de uma interface bem denida sep-
arando a regio turbulenta (T) da regio no-turbulenta (NT). Os processos fsicos presentes nesta
interface so cruciais para o arrastamento turbulento (TE), determinando, e.g., , taxas de crescimento
e mistura em camadas de corte.
O estudo da interface T/NT teve um novo impulso nos ltimos dez anos, com resultados experimen-
tais e numricos questionando o status quo: por exemplo, mostrando que o TE maioritariamente um
processo viscoso das pequenas escalas (nibbling), contrariamente teoria clssica de envolvendo as
grandes escalas (engulfment ). Esta tese avana o conhecimento na rea estudando as estruturas de
vorticidade coerente junto interface T/NT.
As caractersticas geomtricas e dinmicas das estruturas coerentes foram estudados junto inter-
face T/NT. Os resultados obtidos lanam lanam uma luz nova sobre o nibbling dos turbilhes junto
interface T/NT em jactos.
De forma a possibilitar a prossecuo do trabalho realizado foi paralelizado um cdigo de simulao
directa numrica (DNS) espacial de jactos planos turbulentos e duas simulaes massivas realizadas.
Palavras-chave: Jactos Turbulentos, Arrastamento Turbulento, Estruturas Coerentes, Simu-
lao Numrica Directa, Computao de Alto Desempenho, Computao Paralela
Acknowledgements
Like a turbulent ow, the years spent in laboration to deliver this work have been full of interactions,
with a high rate of growth, either scientically, either as a learning process about knowledge itself and
human relations. The global gradient was thus positive and enriching.
Several persons have contributed, in no small way, towards this result, with their support, encourage-
ment, remarks and criticism. To all of them I am deeply grateful. They were instrumental to my hability
of extracting from the surrounding ow the energy and momentum employed to overcome the friction
natural to this kind of endeveour. I will, inevitably, fall in the disgrace of omiting some names by sheer
attrition of my brain cells. To those I ask forgiveness and understanding, this words of aknowledgement
are writen in the nal rush towards the end.
The rst word must be casted in the direction of Prof. Jos Pereira. To him I owe the existance of
the very opportunity to make this PhD into a reality. He has been a friendly lighthouse in a sometimes
rough sea.
Whithout the guidance and help of Doctor Carlos da Silva I have my doubts the ship would have
been steered so well to his nal port. It was a very fruitful cooperation and I can only desire for more to
come. To him I raise a glass at the sound of Beethoven 9th.
I have also a grateful word towards Doctor Nelson Marques, who was in the begginning of this
voyage. To him I owe some important advices some of which I wish now to have followed and always
regard as an example of engineering thought.
A personal word of thanks to Jos Chaves with whom I shared thougths about implementations,
LASEF and CFD in general. I thank his insights and friendship.
To Prof. Sousa also, who was been prompt to answer my comments on the most various matters, a
word of thanks.
To my other collegues in LASEF, to whom I sometimes gnashed my teeth in anger, but with whom I
also shared some laughs and good moments. A special mention should go to Diogo Chambel, to whom
I wish the best of lucks to nish his PhD.
Prof. Sergei Chumakhov spent a brief time in LASEF and I would like to thank him for his comments.
vi
His incitation to do a major overhaul of the parallel plane jet code, before attempting to nish my thesis,
has resulted in a more robust and efcient research tool and, as such, has certainly enriched this work.
To Mrio Lino, my dearest friend since our rst year at IST in Aerospace engineering and who,
besides bestowing me his friendship, allowed me to access much needed computational resources.
This enabled me to process some of the images that can be seen on this pages. To his wife Elsa I
extend my salut and wish for her to quickly and successfully reach the end of her own PhD road.
A small word must also go to Elisabete and Rodrigo: Finally its done after all my missed deadlines.
Thanks for your support and inquiries.
To my adopted family of Rdio Zero who have always been around and managed to hold the fort
when my PhD work demanded me to go away. They are a superb bunch of fellas who ll me up with
pride, and with whom I will always want to share my table. A good word should also for my other radio
friends, Knut, Sarah, Jim, Diana, Pitt, Jay and Paulo who were intrigued by this other part of my life and
always vowed for my success.
To my parents, my brother and his wife (both of which will soon, hopefully, make me an uncle), who
have worried all these years about the nal acomplishement of this task. I hope that by closing this
chapter of my life they can nd some peace. Their concerns are a testimony of their support and love,
to which I shall always be grateful.
Last but never least, to my dear wife
, who silently kicked me forward and gave me the last breath

needed to overcome this challenge and to which I dedicate this document.
On a larger level, two more acknowlegments should be made before proceeding: the rst to the free
software community. It is the wonderful working environment created by their effort that made this work
possible. The second, to the Fundao para a Cincia e Tecnologia, that has nanced this work through
PhD grant BD-24960/2005. To it I thank the trust placed on my use of the Portuguese tax payers money.
Status not recognized by Republic or Church.

vii
ubiquitous Turbulence,
from our blood
to the stars, shapes and enables life.
(which is to say,
we are all sons of chaos)
viii
ix
to C.C.
Table of Contents
Abstract i
Resumo iii
Acknowledgements v
Table of Contents xiii
List of Figures xvii
List of Tables xix
List of Acronyms xxi
1 General Introduction 1
Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
Thesis Organisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
I Background
2 The physics of turbulent ows: a review 7
2.1 Turbulence: general overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Turbulent jets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.3 Turbulent entrainment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
II Numerical tools
3 Direct Numerical Simulations 37
xii TABLE OF CONTENTS
3.1 Context and historical review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.2 Governing Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.3 Computational domain and boundary conditions . . . . . . . . . . . . . . . . . . . . . . . 41
3.4 Spatial discretisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.5 Temporal discretisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.6 Boundary conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4 Parallel plane jet 55
4.1 Parallel architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
4.2 Parallel DNS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4.3 Data dependency analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4.4 Parallel domain decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.5 FFT calculation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4.6 Communications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4.7 Input/Output (I/O) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4.8 Parallel performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.9 Final notes on porting sequential codes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
5 Post-processing tools 71
5.1 Post-processing tools, a foreword . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.2 Vortex identication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5.3 Interface detection and statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
III Results and Discussion
6 The role of coherent vortices near the turbulent/non-turbulent (T/NT) interface in a plane
jet 79
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
6.2 Direct numerical simulation of a turbulent plane jet . . . . . . . . . . . . . . . . . . . . . . 81
6.3 The role of coherent vortices near the turbulent/non-turbulent (T/NT) interface . . . . . . . 87
6.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
7 Intense vorticity structures intense vorticity structures (IVS) near the turbulent/non-turbulent
(T/NT) interface 95
7.1 Direct numerical simulation of a turbulent plane jet . . . . . . . . . . . . . . . . . . . . . . 95
7.2 The intense vorticity structures inside the shear layer region . . . . . . . . . . . . . . . . . 96
7.3 The IVS near the T/NT interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
7.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
TABLE OF CONTENTS xiii
8 Parallel plane jet simulation 115
8.1 Simulations parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
8.2 Parallel performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
8.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
9 Conclusion 127
9.1 Main Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
Bibliography 131
List of Figures
2.1 Turbulent ows of different magnitudes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2 Example of art inspired by turbulence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3 Correlation function sketch . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.4 Longitudinal (f) and lateral (g) velocity correlation functions shape in isotropic turbulence. 11
2.5 Characterisation of energy distribution at different ow scales . . . . . . . . . . . . . . . . 13
2.6 One-dimensional longitudinal velocity spectra . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.7 Smoke plume . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.8 Vortex tubes and sheets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.9 Burgers vortex . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.10 Burgers vortex sheet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.11 Jets in real life . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.12 Jet half-width in the turbulent jet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.13 Plane jet scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.14 Initial plane jet section (vorticity norm ([[) iso-surfaces). Kelvin-Helmholtz velocity prole. 26
2.15 Varicose and sinuous modes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.16 Kelvin-Helmholtz cloud formation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.17 Preferred mode Strouhal number vs. char. length scale . . . . . . . . . . . . . . . . . . . 28
2.18 Intermittency in turbulent ows . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.19 Sketch of entrainment by engulfment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.20 Turbulent plane jet turbulent/non-turbulent interface . . . . . . . . . . . . . . . . . . . . . . 31
3.1 Comparison of temporal and spatial simulations . . . . . . . . . . . . . . . . . . . . . . . . 41
3.2 Computational domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.3 Lateral boundary condition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.4 Coherent structures outow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
xvi LIST OF FIGURES
4.1 Example of cluster layout . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
4.2 Slab division (streamwise planes). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.3 Y 0Z slices division (normal/spanwise planes). . . . . . . . . . . . . . . . . . . . . . . . . 60
4.4 Slab and slices mixed decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.5 IO - Master writes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4.6 IO - All write . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
4.7 IO - MPI-IO . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
4.8 MPI-IO, dedicated I/O nodes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.9 Amdahls law . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.10 Gustafsons law . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.1 Vorticity streamlines conditioned for [[ > 60 . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5.2 Vorticity levels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5.3 Vortex ID stencil . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
5.4 IVS educing sketch . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
5.5 Interface diagram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
5.6 Conditional statistics of isolated interest points. . . . . . . . . . . . . . . . . . . . . . . . . 76
5.7 Vorticity proles conditioned to the T/NT interface distance . . . . . . . . . . . . . . . . . 76
6.1 Q and pressure iso-surfaces of temporal plane jet . . . . . . . . . . . . . . . . . . . . . . . 83
6.2 Vorticity proles conditioned to the T/NT interface distance . . . . . . . . . . . . . . . . . 83
6.3 large vortical structures (LVS) and IVS near the T/NT interface . . . . . . . . . . . . . . . 86
6.4 Determination of l
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
6.5 IVS alignment with T/NT interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
6.6 Vorticity contours and viscous dissipation . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
6.7 Conditional enstrophy proles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
7.1 Temporal plane jet volume fraction histograms . . . . . . . . . . . . . . . . . . . . . . . . 96
7.2 Global IVS properties pdfs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
7.3 Coherent structures near the T/NT interface . . . . . . . . . . . . . . . . . . . . . . . . . . 101
7.4 Coherent structures near the T/NT interface . . . . . . . . . . . . . . . . . . . . . . . . . . 102
7.5 Mean conditional proles near the T/NT interface. Vorticity and number of IVS points. . . 105
7.6 Conditional proles of IVS characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
7.7 Conditional proles of diffusion, stretching time-scales and IVS radius vs. Burgers radius 107
7.8 Temporal evolution of the conditional proles of IVS characteristics . . . . . . . . . . . . . 109
7.9 Conditional proles of stretching rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
7.10 IVS and LVS sketch near the T/NT interface . . . . . . . . . . . . . . . . . . . . . . . . . . 112
8.1 Tangent velocity prole . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
8.2 DNS0: velocity streamlines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
8.3 DNS0: and pr iso-contour surfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
LIST OF FIGURES xvii
8.6 DNS0: Q and pr iso-contour surfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
8.7 DNS0 : transition region zoom . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
8.8 Half-width growth and centreline velocity evolution for DNS0 . . . . . . . . . . . . . . . . . 120
8.9 DNS1: velocity streamlines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
8.11 DNS1: Q and pr iso-contour surfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
8.12 Centreline velocity and half-width growth for DNS1 . . . . . . . . . . . . . . . . . . . . . . 123
8.13 Speedup vs. V0.0 implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
8.14 Speedup for V2.0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
List of Tables
7.1 Characteristics of the IVS in different ows and Re . . . . . . . . . . . . . . . . . . . . . . 98
8.1 Parameters of the parallel turbulent plane jet simulations. Re
H
is the Reynolds number
based on the jet slit width, jet momentum thickness (), x is the grid spacing and noise
refers to the noise percentage added to the average velocity prole, as explained in sec.
(3.6.1) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
8.2 Parameters of the simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
8.3 Parallel versions comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
8.4 Parallel versions comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
8.5 Gigabit Ethernet (GigE) vs InniBand (InniBand) . . . . . . . . . . . . . . . . . . . . . . . 126
8.6 IO performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
List of Acronyms
Turbulence
DES Detached Eddy Simulation
DNS Direct Numerical Simulation
dissipation rate
t
eddy-viscosity
E
k
turbulent kinetic energy spectrum
f
0
instability frequency
f longitudinal velocity correlation function
g lateral velocity correlation function
0.5
jet half-width
HIT homogeneous isotropic turbulence
intermittency factor
IVS intense vorticity structures
Kolmogorov micro-scale
f
c
jet-column mode frequency
jet momentum thickness
H jet inlet height
L
11
integral scale
LES Large Eddy Simulation
LVS large vortical structures
p pressure
RANS Reynolds-Averaged Navier-Stokes
Re Reynolds number
Re
Taylor microscale Reynolds number,Re
=
density
Q positive Q criterion
Q
ij
velocity correlation function
xxii List of Acronyms
shear layer thickness
Str Strouhal number
S
ij
strain rate tensor
L streamwise length
t time
Taylor micro-scale
turbulent/non-turbulent interface vorticity jump thickness

T/NT turbulent/non-turbulent
viscosity
u velocity
u velocity-gradient tensor
vorticity
[[ vorticity norm
uctuating vorticity
Computing
COTS commercial off-the-shelf
CPU central processing unit
FFTW Fastest Fourier Transform in the West
GigE Gigabit Ethernet
HPC High Performance Computing
HTC High Throughput Computing
INGRID Iniciativa Nacional Grid
InniBand InniBand
I/O Input/Output
MPI Message Passing Interface
MPI-IO Message Passing Interface Input/Output
N number of processors
NFS Networked File System
OO Object Oriented
OpenMP Open Multi-Processing
P parallel fraction
SDR Single Data Rate
SMP Symmetric Multiprocessing
S speedup
SS scaled speedup
TB terabyte
xxiii
Numerical
FFT fast Fourier transform
PPT parallel pipelined Thomas algorithm
Other
LASEF Laboratrio de Energia e Fluidos
NASA National Air and Space Agency
xxiv List of Acronyms
CHAPTER 1
General Introduction
This chapter is intended to give in few words a general idea of the subject matter of the work presented,
including its motivation and objectives. Finally, a short description of the document structure is given for
ease of navigation.
Motivation
Turbulence is an ubiquitous state for ows found in nature and/or human generated contexts. More
to it, it enables or amplies a series of physical processes essential to life. As luck would have it, it is
also one of the least understood of physical phenomena, a state of affairs with serious implications in
our everyday life.
Turbulent ows evolving in ambient contexts display a high growth capacity, i.e., to incorporate uid
into them, making it also turbulent. This process is known as turbulent entrainment and is responsible for
the turbulent ow high mixing capacity of mass, energy and scalars (e.g., heat, particles, etc). Turbulent
entrainment is then determinant in jets, ames, plumes, pollutant dispersion, chemical mixing, etc. It is
then clear that being able to correctly understand and model these phenomena is essential to the every
day work of a great number of engineers throughout the world, and impacts the everyday life of the
anonymous citizen. In any engineering area the uncertainty associated with lack of knowledge must be
compensated by safety factors. These induce inefciencies in industrial processes and mechanisms and
so are responsible for economical, and correspondingly, social loss. Increasing our scientic knowledge
about turbulence is then paramount to enable a more efcient use of our planet and propel humanity
towards a sustainable future.
Although ideas about the nature and mechanisms of turbulent entrainment have come to the fore
since the 1950s, lack of detailed information about the region where turbulent entrainment happens,
2 General Introduction
i.e., the area of conuence of the turbulent ow with the irrotational, non-turbulent one, has hampered
their conrmation and validation. New measurements and conditional statistics conducted in this region,
either using data from physical or from numerical experiments, have shed a new light on the problem in
the beginning of the XXI century. In particular, they have shown that one of the most widely accepted
ideas about turbulent entrainment, i.e., that its main driving mechanism would be the engulfment of
irrotational ow by the biggest eddies is not correct. In fact, engulfment accounts only for 10% of the
irrotational mass entrained into the turbulent region. What is the right answer to turbulent entrainment is
still open to debate, but the strongest candidate for the moment lies in the ideas rst ventured by Corrsin
and Kistler in the mid 1950s: i.e., that the main role is played by small eddies motions in a mechanism
known as nibbling.
The above frames one of the purposes of the work presented in this thesis, to contribute to a bet-
ter understanding of turbulent entrainment. To achieve this, attention is given to the role of coherent
vortical structures. These are known to govern the mechanisms of transport, mixing and diffusion of
mass, momentum and scalars amid turbulent ows, and so they seem to play a primary role in the prob-
lem at hand. On focus is the interplay between these structures and the interface that exists between
the turbulent and the non-turbulent regions, the turbulent/non-turbulent (T/NT) interface. The physical
mechanisms and dynamics present at this interface are completely intertwined with the explanation for
the turbulent entrainment.
The study proposed above was enabled by the use of sophisticated numerical tools, able to simulate,
using computers, all the intricacies of the ow. In this respect, the results obtained in the digital medium
are akin to those of its physical counterpart, and classify themselves as full, valid, experiments. Devel-
opments on this type of computational simulations, known as DNS, have been crucial to our increasing
knowledge about turbulence. DNS simulations resolve the full equations describing the behaviour of uid
ows (i.e., the Navier-Stokes equations) and the computer medium allows access to more information,
without the difculties associated with classical experimental instrumentation. The eye of the scien-
tist/engineer is only hampered by its imagination and resourcefulness. Unfortunately, theres no rose
without a thorn: limitations in available memory and computing power impose restrictions in the geome-
tries and range of ows that are achievable in these simulations. In this regard, there is a need to develop
computer codes able to explore new computational architectures and devices, so to keep increasing the
boundaries of what is attainable. A long way has been trodden since the rst DNS simulation of Orszag
in the early 1970s, but growth has been painfully slow. Although evolution in computational power has
increased enormously, the range of scales present in turbulent ows sports an even higher appetite for
resources. To cope with this recourse to the old strategy of joining forces to get stronger is a must:
clusters join large numbers of small CPUs into a tapestry of dedicated networking. This allows access
to higher computational capabilities, either of memory, computing power and storage.
The development of computational tools to take advantage of computational parallel architectures
and so enhance the study of turbulent ows fuels the last motivation for the work here presented.
3
Objectives
The following item list enunciates the objectives of this thesis work:
Implement a pseudo-spectral Direct Numerical Simulation (DNS) numerical code for simulation of
turbulent plane jets in a parallel architecture.
Implement a tracking procedure to study the Intense Vorticity Structures (IVS) in turbulent ows.
Characterise the dynamics and topology of intense vorticity structures (IVS) present in turbulent
plane jets, with the aim to shed light on the Turbulent entrainment mechanism;
Study the interaction between coherent structures and the turbulent/non-turbulent (T/NT) interface;
Realisation of big parallel DNS simulation of a turbulent plane jet.
Thesis Organisation
This document is broken into three parts as follows: the rst is dedicated to layout some notions
about turbulent ows and in particular, turbulent plane jets. Attention is also given to the coherent
vortical structures present in turbulent ows and to the question of turbulent entrainment (chapter 2).
The second part has three chapters. The rst starts with an historical summary of numerical simula-
tion of turbulent ows, especially that of DNS simulation of turbulent jets, and then proceeds to describe
the pseudo-spectral algorithm implemented. Chapter 4 concentrates on the questions surrounding the
parallelisation of the turbulent plane jet DNS code. The next and last chapter of this part deals with
post-processing tools, especially the implementation of the vortex tracking algorithm.
The last part covers the results obtained in the course of this work. These present the results pub-
lished in da Silva and dos Reis (2011) in chapter 6 and those in da Silva et al. (in press) in chapter
7. Chapter 8 presents the results of simulations made with the parallel turbulent plane jet code imple-
mented. To close, some nal thoughts and comments are made.
Part I
Background
CHAPTER 2
The physics of turbulent ows: a review
This chapter provides a summary context of the physical background for the work developed. It begins
with a description about the nature of turbulent ows and their importance, including notions on the
different scales present in these ows and about the energy cascade concept. Coherent structures are
then described, including a short review of eduction methods. The chapter closes with a short review
on turbulent entrainment and the turbulent/non-turbulent (T/NT) interface.
2.1 Turbulence: general overview
Turbulence remains one of the most challenging scientic problems, eluding even a clear-cut, con-
sensual denition (Lesieur (1997); Davidson (2004)). To give an idea of the difculty of the matter it
is enough to quote the not very scientic denition of turbulence (but one to which many would nod in
understanding) due to Bradshaw (1994): invention of the Devil on the seventh day of creation. For
working purposes, and in a more scientic vein, the tentative denition by Kennedy and Corrsin (1961)
(via Davidson (2004)) can be quoted and will set the basis of this small introduction to the physics of
turbulent ows:
Incompressible hydrodynamic turbulence is a spatially complex distribution of vorticity
which advects itself in a chaotic manner in accordance with the vorticity equation. The vor-
ticity eld is random in both time and space, and exhibits a wide and continuous distribution
of length and time scales.
Vorticity () is dened by eq. (2.1):
= u (2.1)
8 The physics of turbulent ows: a review
where u is the velocity vector. In incompressible ows, vorticity, unlike velocity (u), cannot be created
or destroyed in the interior of a uid (by the Kelvin-Helmholtz theorem) and is instead generated by the
presence of boundaries which introduce into the ow the necessary velocity gradients, e.g., the presence
of walls (due to the non-slip condition). Vorticity advects itself through the ow according to eq. (2.2)
D
Dt
= ( ) u +
2
(2.2)
where is the uid viscosity and t the time.
Although the presence of vorticity is essential to the existence of turbulent ow, it is not sufcient: as
expressed by Corrsin, the vorticity eld must be random in both time and space and show continuous
distribution of length and time scales. This is accomplished by the non-linear nature of the Navier-
Stokes equations (eq. (2.3), eq. (2.4))
that describe the ow behaviour of Newtonian uids under the

continuum hypotheses.
Du
Dt
=
u
t
+ (u ) u =
1
p +
2
u (2.3)
u = 0 (2.4)
In eq. (2.3) and eq. (2.4) is the uid density and p the pressure. Newtonian uids are those that obey
Newtons law of viscosity, expressed by
ij
= (u
i
/x
j
+ u
j
/x
i
) (2.5)
where
ij
is the stress tensor of the uid elements.
The main parameter dictating the presence of turbulence in a ow is the Reynolds number (Re),
expressed by
Re =
Ul
(2.6)
where U and l are a characteristic velocity and length of the ow, respectively. This non-dimensional
quantity was rst enunciated by Reynolds (1883) and translates a balance between the viscous and
inviscid forces in the ow and so expresses the ow capacity to dampen (through viscous effects) or am-
plify disturbances present in it. Since most uids of interest have very low viscosity (e.g., air), turbulence
is the norm and not the exception in real life.
Turbulence is ingrained on our universe, from large to small phenomena. It is, in fact, crucial to life
as known to humanity: it is agreed that it is the turbulent convection currents inside the liquid Earth core
that generate the Earth magnetic eld. Without it the planet would be fustigated by the high energy
particles of the solar winds, ejected after by massive solar ares events which are, on the other hand,
caused by the turbulence present in the Sun. On our own planet, again turbulence is found in the small
and the large: the atmosphere can be thought of a giant boundary layer that redistributes water and
heat around the planet, its efciency and dynamics directly linked with its turbulent state. The ocean,
through similar processes is also part of the big mixing system that redistributes heat around our planet.
In gliders, abrasive strips are glued on the wings to trigger transition to turbulence and stop the ow from
Navier-Stokes equations actually refer to the Navier-Stokes and the Mass Conservation equation. They are generally lumped
for easiness of reference.
2.1 Turbulence: general overview 9
separating from the wing and preventing a drastic loss of lift. Inside combustion engines, turbulence
plays a key role in the mixing and efcient burn of fuel. A partial pictorial vision of the large range of
length scales where turbulence can be accounted is given in g. (2.1).
Figure 2.1: Turbulent ows of different magnitudes. From top left, in clockwise rotation: solar coronal ejection, size
of several Earth diameters (source NASA), atmospheric wake caused by Madeira island which has a length of
57Km (source NASA), North-American T6 aircraft, wingspan around 13 meter, doing aerobatics (source Andy
Ciorda@ickr), The Vortex, around 6 meter high, at the Fire Arts Festival in Oakland (source
artmontersgirl@ickr) and blue ink streak, of a few centimetres, mixing with water (source TristanL@ickr)
Besides its importance as a physical phenomenon, with wide consequences to human presence in
the Universe and humans ability to manipulate it, turbulence also has the property to produce some of
the most beautiful events witnessed by man, fascinating and inspiring artists since the beginning of time
and leaving a lasting footprint in artistic pieces worldwide (for instance, in g. (2.2)).
Figure 2.2: Example of art inspired by turbulence: Whirlpools at Naruto, Awa Province, by Hiroshigue, XIX Century
2.1.1 The scales of turbulent ow and the turbulent energy cascade
Big whorls have little whorls
That feed on their velocity,
And little whorls have lesser whorls
And so on to viscosity
(in the molecular sense)
This was how Richardson (Richardson (1922)) rst exposed the idea about how the turbulent kinetic
energy was being handled in turbulent ows. The energy is extracted from the mean ow by the largest
eddies
which, through processes of vortex stretching and breakup, transmit their energy on to smaller
and smaller eddies. The energy transmission process is then mostly inviscid until the eddies reach a size
where viscous effects became dominant (i.e., Re = O(1)) and the energy is dissipated. This concept is
known as the energy cascade, a term coined by Onsager in the 1949.
Because the energy transmission process is essentially inviscid (under the energy cascade hypoth-
esis), the amount of energy extracted from the mean ow must equal the one dissipated at the viscous
end of the cascade. With this in mind some relations can be found to express the energy transfer and
relate the different scales present in the ow.
A very useful mathematical tool to determine the kinematics of turbulence is the velocity correlation
function (Q
ij
), eq. (2.7):
Q
ij
=
i
(x)u
j
(x +r)
_
(2.7)
where the i, j indexes select the velocity components and r the vector between the two space locations of
interest: see g. (2.3). Another important term is u
, which represents the average velocity uctuations

of the ow, obtained from the decomposition of the instantaneous velocity into an average component
For working purposes an eddy can be dened as a velocity difference u across a distance l (Jimnez (2004)).The subject of
what exactly is an eddy is much more muddy than this and no complete agreement can be found between researchers of the eld.
Figure 2.3: Correlation function sketch. The correlation function in eq. (2.7) measures the velocity components
correlation between two points separated by a distance r. Curved lines represent velocity streamlines.
and a uctuating component according to
u = +u
(2.8)
The two most used forms of eq. (2.7) are the longitudinal velocity correlation function (f) and the
lateral velocity correlation function (g). In isotropic turbulence
these assume the form:

Q
11
(re
1
) = u
2
f(r) (2.9)
Q
22
(re
1
) = u
2
g(r) (2.10)
where u
2
veries
u
2
====
1
3
 (2.11)
The two structure functions f,g are dimensionless and satisfy f(0) = g(0) = 1. In isotropic turbulence
they are related by g = f +
1
2
r

r
f and their shape can be surmised from g. (2.4).
Figure 2.4: Longitudinal (f) and lateral (g) velocity correlation functions shape in isotropic turbulence.
The longitudinal velocity correlation function (f) can be used to dene the integral scale, e.g., the
characteristic average length of the largest eddies in the ow, represented by L
11
. This is accomplished
by eq. (2.12)
L
11
=
_

0
f(r)dr (2.12)
In isotropic turbulence the ow properties do not exhibit any preferred direction.

A eddy life time-scale, valid across the different scales, can be obtained relating their characteristic
length and velocity, l and u, respectively:
l
= l/u (2.13)
and from eq. (2.13) an expression for the energy transfer rate can be obtained in the form of eq. (2.14):

u
2
l/u

u
3
l
(2.14)
Since the energy transfer is essentially an inviscid process, until the dissipation region is attained, eq.
(2.14) should also represent the average dissipation rate at the viscous scale u
3
/l. On the
other hand, the dissipation rate , expressed by eq. (2.15),
= 2 < S
ij
S
ij
> v
2
/
2
(2.15)
is connected to the strain rate tensor (S
ij
) (eq. (2.16)):
S
ij
=
1
2
(u
i
/x
j
+ u
j
/x
i
) (2.16)
The u
i
/x
j
+ u
j
/x
i
part in eq. (2.16) belongs to the Newtons law of viscosity stated in eq. (2.5)
on the beginning of this chapter
. Also in eq. (2.15), represents the characteristic size of eddies at

the viscous scale size, v their velocity and
their characteristic time scale. This leads to the relations

expressed by eq. (2.17), eq. (2.18) and eq. (2.19).
(
3
/)
1/4
(2.17)
v ()
1/4
(2.18)
(/)
1/2
(2.19)
Between the integral scale and the Kolmogorov micro-scale there is an intermediary scale worth of
mention. The Taylor micro-scale () is dened by
2
=
u
2
_
_
u
x
_
2
_ (2.20)
which, in homogeneous isotropic turbulence, can be transformed into
2
= 15u
2
/ (2.21)
The Taylor micro-scale lacks a clear physical meaning but it is related with several physical events found
in turbulence. This is evident, for instance, in the results presented in this thesis.
Kolmogorov, in Kolmogorov (1941), added to Richardsons energy cascade picture ( is known as
the Kolmogorov micro-scale) suggesting that at sufciently high Re, the eddies break down process
causes them to loose their anisotropy and become, at the smallest scales, statistically isotropic. Under
this hypothesis, a length scale can be introduced to separate the anisotropic, energy containing eddies,
from the isotropic, smaller ones: l > l
EI
.
Using the strain rate tensor, another expression for Newtons law of viscosity is
ij
= 2S
ij
Kolmogorov built further upon his rst hypothesis venturing that also at high enough Re the statistics
of the small scale eddies (l < l
EI
) have an universal form, dependent only on and . The region
so dened is named universal equilibrium range. A third hypothesis creates a further division in the
universal equilibrium range region by identifying a zone (the dissipation sub-range) where the statistics
are only dependent on and not on . l
DI
will be used to separate these two sub-ranges: the rst
is the inertial sub-range, where viscous effects are negligible and inertial effects predominate, and the
dissipation sub-range, where viscous effects dominate. Fig. (2.5) summarises the scales relations under
Kolmogorovs hypothesis.
Figure 2.5: Characterisation of energy distribution at different ow scales at high Re under Kolmogorov hypothesis.
Flow scales are expressed as Log of the wave number, k = 1/l. L is the ow characteristic length, l
0
the
characteristic size of the largest eddies and the Kolmogorov scale. l
EI
marks the beginning of the inertial
sub-range and l
DI
of the dissipation sub-range.
Obukhov (1941) formulated Kolmogorov hypothesis in the spectral space. In this space the eddies
characteristic length scale l becomes a wave number according to k = 1/l. The energy transfer is
expressed in the spectral space by the turbulent kinetic energy spectrum (E
k
), which gives information
about the energy content at each wave number, i.e., how energy is distributed throughout the wave
numbers. In isotropic turbulence this is expressed by (eq. (2.22)):
E(k) =
2
_

0
R(r)kr sin(kr)dr (2.22)
In eq. (2.22) is R(r) = 1/2 u(x) u(x +r) and r is, as in the correlation function, the distance
between two points in the ow. In the same manner E
k
gauges the energy distribution, its integration
gives the total turbulent kinetic energy when r 0:
1
2
=
_

0
E(k)dk (2.23)
Obukhov also derived from Kolmogorov hypothesis the famous ve thirds power law expressing the
ratio of energy transfer in the inertial sub-range, k
EI
< k < k
DI
, (eq. (2.24))
E(k) = C
k
2/3
k
5/3
(2.24)
where C
k
is the universal Kolmogorov constant, C
k
1.5.
Figure 2.6: Measurements of one-dimensional longitudinal velocity spectra (symbols) and model spectra for
Re
= 30, 70, 130, 300, 600, 1500. Experimental data culled from Saddoughi and Veeravalli (1994), Re
is the last
number in the legend (via Pope (2000)).
Although ample evidence from experimental and numerical data has given support to eq. (2.24) (see
g. (2.6)), some issues still stand, namely what is considered high enough Re. No criterion is given
for this and even in what could be regarded as reasonable high Re ows obtained in laboratory (e.g.,
Re 10
5
) the motions at the dissipative scales are found to be anisotropic (George and Hussein (1991)).
This places some burden upon the assumption of the Kolmogorov universality hypothesis, e.g., the
smallest scales dynamics being independent from the initial conditions. It must be said, nevertheless,
that although important, these discrepancies are expected to have a minor impact on the modelling of
turbulent ows: the small scales (l < l
EI
) contain little energy and anisotropy and so their direct effect
on the ow is reduced. The energy cascade idea, although still a topic of open (and sometimes hot)
discussion, has so far proven to be one of the most useful thought tools in the study of turbulence.
2.1.2 Coherent structures
When observing the evolution of a turbulent ow, e.g., the smoke plume from some burning material
like the one in g. (2.7), it immediately strikes the eye that several three dimensional shapes materialise,
grow and fade in time and space in a somehow choreographed manner. This complex dance is also one
of the mental pictures that most readily jumps to mind when someone speaks of turbulent ows.
Figure 2.7: Smoke plume (credit JOE-3PO@ickr.com)
The eddies turned visible by the smoke are part of what is called coherent structures, a general,
loose term, denominating lumps of uid which form a recognisable entity that can be extracted from its
surroundings.
Several authors provide slightly different denitions. For instance Lesieur (1997) says that coherent
structures will just be dened as a region of space which, at a given time, has some kind of organisa-
tion regarding any quantity related to the ow (velocity, vorticity, pressure, density, temperature, etc.)..
Others authors, like Jeong and Hussain (1995), restrain them to a connected turbulent uid mass with
instantaneously phase-correlated vorticity over its spatial extent. Davidson (2004), on the other hand,
brings also to the fore the idea of preservation in time with "vortical structures which are robust in the
sense that they retain their identity for many eddy-turn-over-times and which appear again and again in
more or less the same form". Sagaut (1998), to close, prefers to keep the subject of temporal preserva-
tion at bay and only considers instantaneous spatial coherence. Regardless of these discrepancies their
presence has attracted the attention of researchers: these structures play an important role in the midst
of turbulent ow and simple mathematical models can be developed to understand their dynamics. They
are also known to govern the transport, mixing and diffusion of mass, momentum and scalars, making
their study essential to engineering applications.
The presence of coherence structures in isotropic turbulence was rst shown by the numerical ex-
periments of Siggia (1981) and conrmed by the experimental observations of Douady et al. (1991).
Lesieur (1997) provides a narrower denition for a subgroup of these coherence structures, the coher-
ent vortices. These are dened as those regions of space that:
Vorticity concentrates enough so that uid trajectories can wind around;
Keep (following the ow motion) a characteristic shape during a life time Tc longer enough in front
of their local turnover time
1
;
Are unpredictable.
Research on coherent vortices has focused both in the eduction of the coherent vortices from the
ow, either from instantaneous elds, either from tracking their evolution in time, and in the study of
their dynamics, especially in the light of analytical solutions of the Navier-Stokes equations for isolated
vortices. Both themes will be further developed below as they are somehow intertwined.
Firstly, concentrating on the question of instantaneous spatial coherence, coherent vortices fall into
two main groups: vortex tubes and vortex sheets. The former consist of elongated tube like structures,
mainly subjected to axial strain, and the latter to vorticity sheets under plain strain (Fig. (2.8)).
Figure 2.8: Two different views of the same region in a homogeneous isotropic turbulence (HIT) simulation using
|| iso-surfaces to show a mixture of coherent tubes and sheets in the ow eld.
Two particular solutions of the Navier-Stokes equations have been useful in the study of coherent
vortices: the Burgers vortex and the Burgers vortex sheet.
Burgers vortex
The Burgers vortex is an exact solution of the Navier-Stokes equations, consisting of a vortex tube
immersed in an axisymmetric, irrotational eld (Sagaut (1998)) with axial vorticity given by
z
(r) =

2
e
r
2
/4
2
(2.25)
where = 2
_
0

z
(r)rdr is the circulation and is time dependent and given by eq. (2.26)
2
(t) = / +
_
2
0
/
_
e
t
(2.26)
In eq. (2.26), t is time,
0
= (t = 0) and is the time independent rate of strain which, for the axial
strain case, is =
0
=
0
T
.S.
0
/[
0
[
2
. The irrotational velocity eld, in cylindrical coordinates, is given
by eq. (2.27):
u
z
= 2z u
r
= r u
=

2r
_
1 e
r
2
/4
2
_
(2.27)
For t large enough,
2
= /, and an equilibrium solution is found for the Burgers vortex, where the
viscous diffusion of vorticity is balanced by the inward production due to axial straining (see g. (2.9)).
Figure 2.9: Burgers vortex in equilibrium. The inward vorticity production balances the outward viscous vorticity
diffusion while the vortex is submitted to axial straining by the background vorticity eld.
The induced kinetic-energy dissipation eld is given by eq. (2.28) and, for the equilibrium solution, it
is independent of . Its peak is also proportional to
2
/
4
and vanishes outside a circular area of order
O(
2
) (Sagaut (1998)).
= 12
2
+

2
16
2
4
_
e
r
2
/4
4
1 e
r
2
/4
4
r
2
/4
4
_
(2.28)
Burgers vortex sheet
Another solution of interest is that of the Burgers vortex sheet, resulting from the combination of a
plane potential ow with a plane shear layer (see g. (2.10)): it consists of a diffusing vortex sheet with
stretched vortex lines (Sagaut (1998)).
The vorticity equation for the Burgers vortex sheet is expressed by eq. (2.29) and the components
of the potential velocity eld by eq. (2.30).
z
=
4
e
y
2
/
2
(2.29)
Figure 2.10: Burgers vortex sheet. Generated by shear, the sheet remains in balance when the compression
caused by the background straining eld balances the viscous diffusion from the vorticity sheet.
u
x
= x u
y
= 0 u
z
= z (2.30)
In eq. (2.29) U is the velocity difference across the shear layer and given by eq. (2.31).
2
=
2
_
1 e
2t
_
(2.31)
The equilibrium solution for the vortex sheet is
2
= 2/ (2.32)
Both these models have been found to provide a good t to coherent vorticity structures found in
turbulent ows. The Burgers vortex will be further discussed in the results chapters of this thesis (see
chapter 6 and 7).
Eduction methods
Eduction methods of coherent vortices are formulated around ideas of isolating certain character-
istics of these structures from the surrounding ow. Some are based on analytical results like those
presented for the Burgers vortex and vortex sheet described previously while others explore results
from the relations of the vorticity vector with the strain eld. The vortex eduction method used on this
work is based in the Burgers vortex model. It was rst presented by Jimenez (1992) and is described
in detail in section 5.2 of chapter 5. For interest and context a review of other educing methods is given
below.
After reordering of the S
ij
eigenvalues into the form
1

2

3
an interesting result emerges: it
has been veried that the vorticity vector aligns itself preferentially with the middle eigenvector, e
2
, of
S
ij
. Jimenez (1992) has shown this alignment with the middle eigenvector happens automatically near a
vortex whose maximum vorticity is large when compared with the background ow vorticity. According to
Jimenez (1992), the background vorticity is dened as that corresponding to <
, where
is dened
by eq. (2.33)
2
= / (2.33)
Another approach relies on the analysis of the velocity-gradient tensor (eq. (2.34)) where
ij
=
(u
i
/x
j
u
j
/x
i
).
u
i
= S
ij
+
ij
(2.34)
The velocity-gradient tensor is explored through the notion that, inside the vortex, vortical effects
should be more pronounced than those effects due to the strain irrotational part. Hunt et al. (1988) was
the rst to propose one such criterion, the positive Q criterion (Q). This criterion denes a region of
space where the second invariant of the velocity-gradient tensor is positive (eq. (2.35)). Q is equivalent
to
2
p > 0 because it has the same sign as the pressure eld Laplacian.
Q =
1
2
(
ij
ij
S
ij
S
ij
) =

2
p
2
(2.35)
Chong et al. (1990) proposed the criterion, where a vortex was identied as a region where eq.
(2.36) holds.
=
_
Q
3
_
2
+
_
det(u)
2
_
2
> 0 (2.36)
An extension of the criterion is the swirling-length criterion. Formulated by Zhou et al. (1999) it
additionally uses the eigenvalues from velocity-gradient tensor (u) as a measure of swirling rate in-
side the vortex. Besides fullling eq. (2.36) it requires that the complex part of the complex-conjugate
eigenvalues of u,
cr
i
ci
, should be above a pre-determinate threshold. Chakraborty et al. (2005)
proposed an enhancement starting from the idea of using a local frame associated with the u eigen-
vectors. They proposed that a vortex would satisfy the relations in eq. (2.37)
ci
and

cr
ci
(2.37)
where , ,
are positive threshold values.

The
2
criterion, from Jeong and Hussain (1995), is also based on ordered eigenvalues of u (
2
being the middle eigenvalue). A vortex, according to this criterion, is a space region where
2
is negative:
2
< 0 (2.38)
Horiuti (2001) expanded on the
2
criterion. He re-ordered the eigenvalues according to:
z
as the
most aligned with vorticity (),
+
as the largest and
the smallest. A vortex is then dened as a

region where eq. (2.39) holds.
0 >
+

(2.39)
This criterion obtains a similar region to that of the Burgers vortex core.
Vortex sheets, on the other hand, have gathered less attention from the community, being also harder
to single out from the ow: they have briefer lives, are less organised and stable and the designation
encompass different objects. In this respect, for instance, Horiuti (2001) makes the distinction between
at sheets, more akin to the Burgers vortex sheet, and curved sheets, that exist surrounding the core of
a vortex tube. Tanaka and Kida (1993) dened at sheets
with the criterion in eq. (2.40)

1
2
<
[W
ij
[
2
[S
ij
[
2
<
4
3
(2.40)
Horiuti (2001) uses the same previous re-ordering to dene at sheets as regions conforming to eq.
(2.41) and curved sheets to those that obey eq. (2.42).
The authors use the terminology strong vortex instead of at sheets.

+
0
(2.41)
> 0 (2.42)
As the criterion proposed in eq. (2.41) has been noticed, in some contexts, to also educe vortex
cores, Horiuti and Takagi (2005) proposed a new criterion, based on the eigenvalues of S
ij
W
ij
+W
ij
S
ij
.
Using a similar eigenvectors ordering for this tensor as before (for the eigenvalue associated with the
eigenvector most aligned with the vorticity (
z
), to the largest (
+
) and smallest (
)) the criterion uses

an arbitrary threshold and the form of eq. (2.43). Besides homogeneous isotropic turbulence (HIT) the
criteria was successfully applied in a homogeneous shear ow to study stretched spiral vortices (Horiuti
and Ozawa (2010)).
+
> (2.43)
2.2 Turbulent jets
2.2.1 Turbulent jets: a context
Turbulent jets are part of the large family of free shear ows i.e., ows developing far away from
the inuence of boundaries (i.e., walls) and usually into a quiescent environment. Other members of this
family are, for instance, wakes and mixing layers. Examples of jets in our surroundings are given in g
2.11 and, as can be seen, the interest in their study is twofold: a direct need of predicting their behaviour
and inuence, stemming from their presence on such a broad range of man made and natural contexts,
and that of being a good object of study to further the knowledge about the physics of turbulent ow.
As the images in g. (2.11) illustrate, a jet consists of a source of momentum that issues into an
ambient environment. For the sake of simplicity, it will be admitted that this environment is at rest or
animated of constant velocity in the same direction of the momentum injection. Jets, like other free
shear ows, evolve and develop along a preferred direction. This points already to some important
features for this kind of ows: their average streamwise velocity, u
x
, is much higher than their transverse
velocity, u
y
(i.e., u
x
>> u
y
), and their transversely gradients /y, much higher than the streamwise
ones, /x (i.e., /y >> /x). Also worthy of note and highly visible in g. (2.11) is the relation of
scales between the shear layer thickness () and the streamwise length (L): the latter being much larger
than the former, i.e., /L << 1.
The examples on g. (2.11) also highlight a striking characteristic common to all shear ows: the
sudden change from a turbulent, rotational region, to an irrotational surrounding. This change is me-
diated by a highly convoluted interface, the turbulent/non-turbulent (T/NT) interface
. It is through this
interface, as shall be described further into the text, that important exchanges of mass, momentum and
scalar (e.g., temperature) take place, while the jet develops in the streamwise direction, spreading out.
It is important to note that the T/NT interface may not necessarily correspond to the visual outline perceived from g. (2.11).
2.2 Turbulent jets 21
Figure 2.11: Jets in real life. Clockwise, from left, diesel engine injector (Credit J.R. Thompson Company), re
hose (Credit omar_infant-ramos@ickr.com), Soyuz TMA-20 rocket (Credit NASA@ickr.com) and Mount
Cleveland erupting, in Alaska (CreditNASAs Marshall Space Flight Center@ickr.com).
This spreading makes the jet swell, imparting momentum on the adjoining uid, contaminating it with
vorticity, and dragging mass onto it in a phenomena known as turbulent entrainment. A more detailed
discussion on the nature of turbulent entrainment will be presented further into the text.
The T/NT interface topology also seems to change around Re 10
4
(Dimotakis (2000)). Below
this threshold the interface surface shape is dictated mainly by the larger structures and little ne-scale
turbulence is seen. Above it, the turbulence appears more developed, with prominence for the ne-scale
type of turbulence. No explanation to this metamorphosis has been posited so far.
Because the T/NT interface has such a convoluted nature it shows great intermittency, making it dif-
cult to ascertain the jet boundaries and, consequently, properly quantify the jet growth. This motivates
the use of a quantity more amenable to be obtained, either experimentally or computationally: the jet
half-width (
0.5
). This last expresses the distance from the jet centreline to the location where the av-
erage streamwise velocity is half of the average centreline streamwise velocity (see eq. (2.44) and g.
(2.12)) for illustration.
u
x
(x, y =
0.5
) = 1/2 [ u
x
(x, y = 0) u
x
(x, y = )] (2.44)
Another important variable to characterise the plane jet is the , a measure of the jet momentum
departure from an ideal (inviscid) situation:
(x) =
_

0
_
u
x
(x) u
(x)
u
0
u
(x)
_ _
1
u
x
(x) u
(x)
u
0
u
(x)
_
dy (2.45)
Figure 2.12: Jet half-width (
0.5
) in the turbulent jet, where U
0.5
comes from eq. (2.44) and U
c
= u
x
.
In eq. (2.45), u
(x) is the undisturbed velocity far away from the jet, u

0
the centreline velocity at the
origin while u
x
(x) is the local centreline velocity.
2.2.2 The turbulent plane jet
General description
The turbulent plane jet designates a jet issuing into the environment through a rectangular slit of
height H (see g. (2.13)) and developing along the XoY plane, i.e., the mean ow on parallel planes is
statistically identical.
Figure 2.13: Plane jet schematic views (side and top-down) showing its different regions. || iso-surfaces are used
for illustration purposes.
In the portrait presented in g. (2.13) several regions can be distinguished. There are two shear
layers, analogue to mixing layers, originated by the high velocity gradient between the ow incoming
from the inlet slit and that of quiescent outside environment. The ow between those shear layers in the
beginning of the jet is called the potential core region. Free from the inuence of the viscosity effects
that created the shear layers, it keeps the initial inlet velocity prole. Other regions worth of mention
include the transition region, encompassing all the region of space where the ow undergoes transition
into a fully developed turbulent state (see sec. 2.2.3) and, later on, further downstream, the so called
self-similarity region, where the velocity proles (and those of some other variables of interest) collapse
seamlessly after being properly scaled (detailed further below).
Governing equations
A set of governing equations can be set for the turbulent plane jet, taking into consideration a group
of simplications particular to such type of ows (Davidson (2004)):
The order of the streamwise gradients of the Reynolds stresses,
R
ij
/x, is much smaller than
those of the transverse gradients
.;
Negligible laminar (viscous) stresses;
Very small normal component of the mean inertia term ( u ) u
y
which is of the order of u
2
/(radius
of curvature of the mean streamlines).
With the above both the streamwise and transverse equations of motion simplify into
( u ) u
x
=

y
_
R
xy
p
x
(2.46)
y
_
R
yy
p
y
= 0 (2.47)
Both eq. (2.46) and eq. (2.47) can be further combined into a single equation after noting a few more
simplications. Namely, as the external ow is uniform or quiescent, the pressure at innite (P
) is
constant and the dependency of p on x is coupled to that of (u
2
y
) through eq. (2.48):
p + (u
2
y
) = p
(x) (2.48)
Because the longitudinal gradients in the Reynolds stresses are negligible, eq. (2.46) becomes
( u ) u
x
=
R
xy
y
(2.49)
Coupling eq. (2.49) with u = 0 a simplied equation for the momentum is obtained:
x
( u
2
x
) +

y
( u
x
u
y
) =
R
xy
y
(2.50)
Knowing that u
x
(y = ) = 0 (absence of co-ow), eq. (2.50), integrated in the transverse direction for
y results in eq. (2.51):
M =
_

u
x
(x, y)
2
dy,
M
x
= 0 (2.51)
The Reynolds stresses arise from the averaging of the Navier-Stokes equation (eq. (2.3)) and are of the form
R
ij
= u
i
u
j
.
They appear to act like stresses in the ow but in fact represent the ux of momentum in or out of a xed volume of uid due to the
turbulent uctuations of velocity.
Eq. (2.51) states that the average momentum ux in a jet is conserved.
Although the average momentum ux is conserved, the average mass ux is not (as already men-
tioned, the jet mass grows in the streamwise direction through turbulent entrainment). This is described
by eq. (2.52):
m(x) =
_

u
x
(x, y)dy,
m
x
> 0 (2.52)
Turbulent jet self-similarity
Through experimental observation (Bradbury (1965); Wygnanski et al. (1976)) there is evidence that
after a certain distance (from x/H 40) the mean-velocity and Reynolds-stress proles become self-
similar after being properly scaled by the centreline velocity u
c
= u
x
(x, y = 0) and jet half-width (
0.5
),
i.e., they collapse along the streamwise direction. This can thus be expressed by
u
x
u
c
(x)
= f(y/
0.5
(x)) = f() (2.53)
With eq. (2.53), the equation for the momentum ux eq. (2.51) and mass ux eq. (2.52) become, for the
self-similar region
M =
_

u
x
(x, y)
2
dy = u
2
c
0.5
_

f
2
()d (2.54)
m =
_

u
x
(x, y)dy = u
c
0.5
_

f()d (2.55)
Furthering into the question it can be shown from eq. (2.54) that
u
2
c
0.5
= constant (2.56)
and an entrainment coefcient can be dened for the jet (Davidson (2004); Pope (2000)):
d
0.5
dx
= = constant (2.57)
For the plane jet it has been found, tting to experimental data, that 0.21, which gives a semi-angle of
6
. Also, from Thomas and Chu (1989); da Silva and Mtais (2002a); Stanley et al. (2002), the following
linear expressions describe the evolution of
0.5
and u
c
in the centreline of the turbulent plane jet:
0.5
H
= K
u1
_
x
H
+ K
u2
_
(2.58)
_
u
c
(x) u(x = 0, y = )
u
c
u(x = 0, y = )
_
2
= C
u1
_
x
H
+ C
u2
_
(2.59)
The constants K
u1
,C
u1
dene the growth ratio and K
u2
,C
u2
the virtual origin of the jet.
Taking into account the Boussinesq-Prandtl eddy-viscosity (
t
) hypothesis
, another relations can

be obtained for the self-similar region of the turbulent plane jet. Assuming
t
to be independent of y and
a function of the local values of u
c
(x) and (x), eq. (2.49) becomes
u
x
u
x
x
+ u
y
u
x
y
= bx u
c
2
u
x
y
2
+ (2.60)
In layman terms the Boussinesq-Prandtl eddy-viscosity hypothesis argues that the effect of turbulent mixing of momentum
R
xy
is analogous to the molecular transport of momentum, leading to a laminar stress
xy
. This would be equivalent to increase the
ow viscosity by a turbulent viscosity term
t
(Davidson (2004)).
where b is the eddy-viscosity coefcient from
t
= b(x) u
c
(x). Eq. (2.60) has a self-similar solution
which, after some manipulation (see Pope (2000) for a detailed derivation), gives the following expres-
sion for the self-similar prole of the turbulent plane jet:
f() = sech
2
() (2.61)
with = 1/2ln(1 +
_
(2))
2
0.88.
Additional discussion on these relations is postponed to chapter 8, where they will be placed in
context with the results obtained during in the course of this work.
2.2.3 Transition to turbulence
Turbulence is the norm, rather than the exception, in real life ows and jets are no exception of this.
After some time, the small disturbances present in the surrounding environment (or already incorporated
in the ow) are amplied and the ow becomes turbulent, a process known as transition. Some common
sources of disturbance are
Free stream disturbances
Inlet perturbations (initial conditions)
Pressure waves from acoustic sources
Mechanical structures vibrations
Transition is, by itself, a broad subject, the particulars of which have a strong ow type dependence.
Some general principles do apply though across the plethora of ow situations, namely the existence of
the afore mentioned disturbances coupled with a ow receptivity to amplify those same disturbances
(which in general implies a high Re).
Two instability length scales (and, consequently, two frequencies) are found to govern the transition
process in the plane jet (Ho and Huerre (1984)): that of the shear layer region, L
inst
= the shear
layer mode and that associated with the region starting after the end of the potential core, L
inst
= H
the preferred or jet-column mode.
2.2.4 Shear layer mode
The initial jet transition is connected to the behaviour of the two shear layers resulting from the step
velocity gradient between the inlet velocity prole and the quiet (or relatively low, uniformvelocity) outside
environment. These shear layers grow through momentum diffusion towards the outside and into the
potential core region (see g. (2.14)).
It is on the shear layers that the two main actors (or most striking features) of this initial phase of
the jet development appear: the big Kelvin-Helmholtz instability. Forming in the upper and lower part of
the jet, they are initially separated by the potential core region, developing in a independent, although
correlated, fashion, depending on the inlet velocity prole (Thomas and Prakash (1991)): at proles
promote a symmetrical arrangement of primary Kelvin-Helmholtz instability the so called varicose
mode, parabolic proles lead to asymmetric arrangements sinuous mode. Flat velocity proles are
Figure 2.14: Initial plane jet section (|| iso-surfaces). Kelvin-Helmholtz velocity prole.
connected with short nozzles with high contraction ratio, as noted, for instance, by Sato (1960) and
Thomas and Goldschmidt (1983). For a graphical differentiation between the two modes see g. (2.15).
Figure 2.15: Varicose and sinuous modes.
Before the end of the potential core, the shear layer mode frequency is associated with the Strouhal
number given by eq. (2.62)
Str
=
f
0
0
U
0
(2.62)
where is the initial jet momentum thickness, f
0
is the instability frequency, f
0
= O(1/
0
) and U
0
the
maximum inlet streamwise velocity. For the plane jet Str
= 0.0165 (see Michalke (1965)).

The Kelvin-Helmholtz vortices result from perturbations in the shear layer velocity prole (see g.
(2.14)). A small oscillation in the area between streamlines (like point A in g. (2.14)) will create an
asymmetry between the streamlines, accelerating the ow where the area is reduced (- sign) and
decelerating it on the opposite side. The velocity increase will cause a pressure decrease, further
amplifying the effect as the streamline is pulled in that direction. The vorticity convection by the mean
ow will then increase the distance between the crests and troughs, shaping the ow into a characteristic
wavy pattern (which can be sometimes admired in the sky, as seen in g. (2.16)...).
Figure 2.16: A good way to visualise the Kelvin-Helmholtz instability is looking into the sky. (Detroit, by Jonathan
Caves@ickr)
2.2.5 Preferred and column mode
After the potential core region, as already remarked, the instability length scale is L
inst
= H, identi-
ed as the preferred or jet-column mode. The relation between the two disparate instability length scales
(and associated frequencies) differentiates between two categories of shear layer adjustment: the tuned
jet has f
0
/f
c
= 2
n
, where n is an integer and f
c
is the jet-column mode frequency, corresponding to
the average passage frequency of vortices at the end of the potential core region. For f
0
/f
c
,= 2
n
the
jet falls into the untuned jet category. This untuned jet mode yields a more dramatic re-adjustment
of frequency and phase from the vortices arriving from the shear layers so they t into the jet column
frequency. To the new instability length scale corresponds a new Strouhal number (Str), given by
Str
H
=
f
c
H
U
0
(2.63)
where f
c
= O(1/H). For this mode, and at high Re, it has been found by several experimental and
numerical works that 0.24 < Str
H
< 0.5 in all jets (for instance, the experimental works of Crow and
Champagne (1971), Yule (1978), Gutmark and Ho (1983), Zaman and Hussain (1980) and Camussi
and Guj (1999) and the numerical works of Olsson and Fuchs (1996), Danaila and Boersma (2000) and
Urbin and Mtais (1997)). Ho and Huerre (1984) studied the variation of the Str for different jets, either
plane or round. Fig. (2.17) exhibits the variation of Str with H/
0
for the plane jet from Ho and Huerre
(1984): it can be seen that after a certain value of H/
0
, the Str stays put at 0.25.
2.2.6 Secondary instabilities
Due to the instability modes varicose or sinuous secondary instabilities will develop from the main
vortex rollers arising from the Kelvin-Helmholtz instabilities (which sometimes also grow through merging
among themselves). Although the characteristic length of these secondary instabilities is lower than that
of the primary ones (da Silva (2001)), they promote an increase in the three dimensional complexity
of the ow (re-check g. (2.14)). This complexity is further magnied beyond the potential core region
Figure 2.17: Plane jet preferred mode Strouhal number vs. characteristic length scale ratio. From Ho and Huerre
(1984).
with the appearing of smaller 3D structures, lacking preferential direction and so characteristic of fully
developed turbulent ow.
2.2.7 Merging of the shear layers
The aforementioned evolution is coupled with the shear layers growth and their inevitable merging
as they are convected downstream. This will cause the engulfment of the potential core region around
x/H 4 5 and originate a complex interaction between them which will climax with a violent restruc-
turing of the ow when they nally merge (Thomas and Prakash (1991), Thomas and Chu (1989)). The
location of the potential core end moves towards the inlet with increasing Re (Deo et al. (2007)).
The merging and interaction of the large scale opposite sign vortices after the potential core region
is, as remarked by Hussain (1986), unique to the plane jet.
After this, if the arrangement of the Kelvin-Helmholtz vortices was symmetrical (from a near top-hat
inlet velocity proles), it quickly changes into an array of self-preserving anti-symmetrical structures that
live through the rest of the similarity region. This transformation is the result of the ow self-adapting
from the preferred mode to the column mode mentioned in sec. (2.2.5). Thomas and Chu (1989) showed
that it is the upstream feedback from these restructuring events that leads to the symmetry loss of the
varicose mode Kelvin-Helmholtz vortices arrangement. For jets with an initial dominance of the sinuous
mode these events are mode mild because the vortices arrangement is already anti-symmetrical. It
follows that after this events the varicose mode, if existent, disappears and sinuous mode dominates
(da Silva and Mtais (2002a)). In direct result of all this a new frequency emerges, named similarity
fundamental mode, characterised by a Str of:
Str
sim
=
f
sim
0.5
U
= 0.11 (2.64)
where
0.5
comes from eq. (2.44) and U = U
1
U
2
, with U
1
the maximum inlet velocity and U
2
the
inlet co-ow velocity.
2.3 Turbulent entrainment 29
2.3 Turbulent entrainment
2.3.1 General context
Turbulent shear ows have, in contrast with their laminar siblings, a high growth capability, contam-
inating large swathes of non-turbulent ow. This has important implications in their surroundings, en-
abling them, for instance, to sustain higher mixing rates of mass, energy and scalars (e.g., , temperature)
than its laminar siblings. This growth process is called entrainment and its physics are well understood
for the laminar regime where it is governed by an essential viscous diffusion process. Unfortunately,
for the turbulent regime, many questions remain unanswered. Several theories have been put forward
from experiences and theoretical studies (Corrsin and Kistler (1955); Phillips (1955); Townsend (1966);
Carruthers and Hunt (1986) but lack of detailed information, e.g., velocity maps of the turbulent/non-
turbulent (T/NT) region, have precluded their conrmation. In this regard the emergence of better ex-
perimental apparatus and new DNS simulations has shed another light on the problem since the early
2000s, especially with studies of Bisset et al. (2002); Mathew and Basu (2002); Westerweel et al. (2005);
Hunt et al. (2008); da Silva and Pereira (2008); da Silva and Taveira (2010).
2.3.2 Engulfment vs. nibbling
It is understood that explanation of the turbulent entrainment phenomena must be sought in the
dynamic interactions found in the interface between the turbulent and the non-turbulent ow regions: the
T/NT interface.
The nature of this interface was rst discerned by visual inspection and later better ascertained by
noting the high intermittency of the signal in probes inserted into the ow to account for the presence of
turbulent ow. First measured by Townsend (1948), this behaviour was quantied by the intermittency
factor () which gives the relative time that a probe spends immersed in turbulent ow. It has also
become clear that the interface separating these two regions (the turbulent and non-turbulent) was highly
convoluted (g. (2.18)).
Corrsin and Kistler (1955) rst elaborated upon the way how irrotational ow becomes part of the
turbulent region and thus ventured a theory to explain the nature of turbulent entrainment. Knowing that
vorticity could not be created on the midst of irrotational ow (a result from the Kelvins theorem), the
idea of a viscous superlayer with a characteristic width of the order of Kolmogorov micro-scale () scale
was proposed. Vorticity would be diffused across this layer into the irrotational ow, thus contaminating
it, in a process akin to that happening in laminar ows.
The acknowledgement of the presence of large-scale organised structures in the ow near the T/NT
interface by Brown and Roshko (1974) furthered a new hypothesis about the nature of turbulent entrain-
ment, expounded later by Townsend (1976). This explanation has become the staple theory to explain
turbulent entrainment, still being the standard found in uid mechanics textbooks worldwide. The en-
gulfment idea due to Townsend argues that turbulent entrainment is mainly driven by the large scale
motions. These envelope packs of irrotational uid into the inside of the turbulent region, which later
become turbulent through a digestion process of molecular diffusion, as described by Corrsin (see g.
Figure 2.18: Several instantaneous snapshots (same region at different times) of bi-dimensional vorticity
iso-contour highlighting the intermittency of the vorticity presence signal in the same locations in space. It can be
seen that the values of the red probe vary in time, indicating intermittency in the signal. Flow from left to right, 1
signals turbulent ow, 0 irrotational ow.
Figure 2.19: Sketch of entrainment by engulfment, inspired by Corcos and Sherman (1984). Turbulent ow in grey,
arrows represent streamline segments and are not to scale.
(2.19)).
Several studies have supported the idea that engulfment is the dominant process for turbulent en-
trainment, e.g., Dham and Dimotakis (1987); Ferr et al. (1990); Dimotakis (2000). New studies, how-
ever, targeting more detailed measurements in the T/NT region have proven that the role of engulfment
in the turbulent entrainment process has been overestimated (Bisset et al. (2002); Mathew and Basu
(2002); Westerweel et al. (2005, 2009)). In these studies conditional statistics with the interface distance
show that the amount of irrotational uid trapped by the large-scale movements only accounts for a small
percentage of the total mass ux into the turbulent ow, e.g., 10%. This has brought once again to
the fore the ideas exposed by Corrsin of the main phenomena driving the turbulent entrainment being
connected with the small scale (e.g., nibbling). Albeit these late development on the knowledge about
entrainment, it is still thought that the entrainment rate is linked to the large scale motion present near
the interface.
2.3.3 T/NT interface and turbulent entrainment
Figure 2.20: Bi-dimensional vorticity iso-contour of a developed turbulent plane jet highlighting the high contorted
nature of the turbulent/non-turbulent interface.
As mentioned, between the turbulent and the non-turbulent region a sharp, convoluted interface can
be found, as detailed rst by Corrsin and Kistler (1955) (see g. (2.20)). This interface marks an area of
change in the nature of the velocity uctuations present in the ow from vortical to irrotational (u = 0).
Phillips (1955) has shown that important velocity uctuations exist outside the turbulent region near the
T/NT interface and estimated a decay law of these velocity uctuations with the distance from the T/NT
interface: for distances higher that integral scale (L
11
), i.e., u
2 y
4
I
[ y
I
/L
11
> 1. da Silva and
dos Reis (2011) have shown that dissipation measured in the irrotational region, near the interface, is
connected with velocity uctuations caused by the coherent structures nearby.
It must be said that the T/NT interface is not the viscous superlayer mentioned by Corrsin but is made
of the adjacent turbulent region near the irrotational region (see Bisset et al. (2002)). The superlayer
would rest on the top of this interface or be part of it, as a last thin belt prior to the irrotational region. Its
existence is still open to debate.
Bisset et al. (2002) observed that the vorticity components undergo a sharp jump at the T/NT in-
terface. Westerweel et al. (2005, 2009) reported jumps in streamwise velocity and passive scalar eld
across this interface, conrmed that engulfment is not the dominating entrainment driving process and
analysed the dynamics of the ow relative to the T/NT interface and investigated eddy viscosity near
that location. Holzner et al. (2007, 2008) analysed the dynamics of enstrophy and strain near a T/NT
interface generated by an oscillating grid. They observed the existence of an intense kinetic energy
dissipation outside the turbulent region near the T/NT interface and that the net effect of viscosity at that
location is to cause an increase of the total enstrophy. Subsequently it was observed by Holzner et al.
(2008); da Silva and Pereira (2007b) that it is the viscous diffusion that is responsible for this positive
viscous contribution to the enstrophy while the enstrophy viscous dissipation remains negative across
the jet, as expected. Recently Holzner et al. (2009) decomposed the entrainment velocity into two terms:
a viscous and an inviscid term. It was observed that the viscous term dominates the entrainment ve-
locity and has a magnitude comparable to that of the local Kolmogorov velocity. The universal small
scale features such as the geometry of the dissipation, the geometry of the straining (or deformation)
of the uid elements, and the behaviour of the invariants of the velocity gradient tensor Q and R
during the entrainment process was studied by da Silva and Pereira (2008, 2009). They observed that
the characteristic tear drop shape in the (R, Q) map is not formed at the T/NT interface, but needs
about one Taylor micro-scale into the turbulent region to form completely. Finally, the challenges faced
by the subgrid-scale models near the T/NT interface were analysed by da Silva (2009), showing that the
classical subgrid-scale models may need particular modications near the edge of a jet if ne capturing
of the Reynolds stresses at this location is needed e.g. in order to predict accurately the mixing rates
near the T/NT interface of a jet.
Conicting results have been reported for the thickness of the sharp vorticity jump (
) observed at
the T/NT interface. Reynolds (1972) predicted that a vorticity jump is to be expected to occur at the
T/NT interface but the observed values for its thickness vary in the literature e.g., the thickness of the
turbulent front generated from an oscillating grid is of the order of the Kolmogorov micro-scale (Holzner
et al. (2007)) whereas the experimental results from round jets by Westerweel et al. (2005, 2009) and
the numerical simulations of plane jets by da Silva and Pereira (2008, 2009) clearly show a vorticity jump
thickness at the T/NT interface of the order of Taylor micro-scale. Finally, da Silva and Taveira (2010)
were able to reconcile these reports showing that
is roughly equal to the size of the largest coherent

vortices found nearby, i.e., <
< .
Many aspects associated with the characteristics and dynamics of the T/NT interface do, however,
remain unknown. One of these is the existence of a viscous super-layer where viscous processes
dominate, as suggested by Corrsin and Kistler (1955).
2.3.4 intense vorticity structures (IVS) vs. large vortical structures (LVS) near the turbulent/non-
turbulent (T/NT) interface
It is useful to divide the coherent vortical structures present in the ow into two types: intense vorticity
structures (IVS) and large vortical structures (LVS). The LVS are the largest vortical structures which
are present in a particular ow. Often originating in the particular instabilities from that ow, their char-
acteristics such as the vortex core radius, length size and life time are deeply related to these processes
and therefore are quite different from ow to ow. However the dynamics of these structures share some
common features with LVS from other turbulent ows e.g., they consist in structures with roughly tubular
shape and are approximately governed by the same simple inviscid laws.
On the other hand the IVS are dened as structures with particularly strong vorticity i.e., structures
comprised of ow points with vorticity greater than a particularly high vorticity threshold,
ivs
. Jimnez
et al. (1993) dened
ivs
as equal to the vorticity dening the 1% of ow points with the h highest
vorticity. In isotropic turbulence the IVS are the well known worms described by Siggia (1981). Many
works have been devoted to the study of these structures, particularly in isotropic turbulence e.g., Siggia
(1981); Ashust et al. (1987); Vincent and Meneguzzi (1991); Jimnez et al. (1993); Jimnez and Wray
(1998), but recently other ows have been used to study the IVS such as mixing layers (Tanahashi et al.
(2001)), channel ows (Kang et al. (2008)), and jets (Ganapathisubramani et al. (2008)). In contrast to
the LVS the characteristics of the IVS are similar in a variety of different ows e.g., the vortex radius,
axial vorticity and azimuthal velocity from the IVS in mixing layers are similar to the values found in
isotropic turbulence. Specically, it has been reported that the radius of the IVS is R 5, where is
the Kolmogorov micro-scale, in isotropic turbulence (Jimnez et al. (1993)), mixing layers (Tanahashi
et al. (2001)), channel ows (Kang et al. (2008)), and jets (Ganapathisubramani et al. (2008)). That the
threshold used to dene the IVS has a negligible impact on their computed characteristics is attested by
the fact that other works use different techniques to dene the worms obtaining very similar statistics
e.g., Kang et al. (2008). In isotropic turbulence Jimnez et al. (1993); Jimnez and Wray (1998) divided
the ow structures into three classes: (i) intense vorticity structures (IVS), (ii) structures of background
vorticity, dened as structures with vorticity greater than the uctuating vorticity and smaller than the
vorticity dening the IVS i.e.,
< <
ivs
, and (iii) weak vorticity structures, with <
. The
uctuating vorticity (
) is given by eq. (2.65):
=
_
<
2
> < >< >
_
1/2
(2.65)
It is important to realise that with the notation used here the LVS are all the ow vortices that are not
included in the set of IVS i.e., the LVS contains also the background vorticity structures and the weak
vorticity structures which in isotropic turbulence are relatively free from vorticity and can be identied
with the velocity eddies.
The important role played by the LVS in the turbulent entrainment has been recognised long time ago
(Townsend (1976)). Past studies assumed that turbulent entrainment is mainly caused by large-scale
engulng motions (Townsend (1976)) induced by the LVS. Bisset et al. (2002) described the topology
of the streamlines induced by LVS near the T/NT interface, and observed that streams of turbulent and
irrotational motion collide at the interface, and stretch out along it, driving its highly convoluted shape.
The vortical interactions with a T/NT interface were also studied in detail using linear analytical models
in Hunt et al. (2008).
Recently, some aspects of the LVS and IVS near the T/NT interface were investigated by da Silva and
Taveira (2010) and da Silva and dos Reis (2011). Specically, it has been shown that in a jet the T/NT
interface is the physical surface dened by the LVS, in agreement with the classical picture that is known
for some time e.g., Townsend (1966); Bisset et al. (2002): the LVS are responsible for the observed
convolutions of the T/NT interface. This explains that these convolutions length scale is roughly equal
to the length scale of the LVS. da Silva and dos Reis (2011) showed moreover that the presence of
LVS is responsible for the existence of a region of irrotational kinetic energy dissipation and by a positive
enstrophy diffusion in the region bounding the turbulent and the non-turbulent ow regions.
The radius of the LVS near the T/NT interface can be estimated bearing in mind that long lived
vortices are vortices where the timescale associated the radial viscous diffusion of vorticity is roughly
balanced by the axial stretching caused by the local strain rate eld S
, such as in a Burgers vortex

(Jimnez and Wray (1998)). The radius of these vortices even if they are not exactly described by
the Burgers vortex model is at least of the order of the Burgers radius R
lvs

_

S
_
1/2
. In a jet
the magnitude of the strain rate acting on a IVS near the jet edge is S
/L
11
, where u
is the
uctuating velocity eld and L
11
is the integral scale of turbulence (Hunt et al. (2010)). This leads to
R
lvs

_

(u
/L
11
)
= L
11
Re
1/2
0
, where Re
0
= u
L
11
/ is the Reynolds number associated with
the integral scale. However, it is possible to nd jets (particularly at very high Reynolds numbers), or
shear free ows, where the biggest existing vorticity structures are similar to the IVS: they exhibit no
particular spatial orientation, due to the fragmentation of the LVS. In this case S
/ and the radius

of the vortices dening the T/NT interface is R
ivs

_

(u
/)
= Re
1/2
, where Re
= u
/ is the
Reynolds number based on the Taylor micro-scale.
Notwithstanding the amount of work carried out on the role of the LVS near the T/NT interface in
free shear ows, there is virtually no work devoted to the effects and dynamics of the IVS in this region.
Specically the dynamics of the IVS near the T/NT interface remains largely unexplored.
Part II
Numerical tools
CHAPTER 3
Direct Numerical Simulations
A small historical review about the numerical simulation of turbulent ows is presented, giving a special
attention to turbulent jets. After this review the pseudo-spectral DNS algorithm for simulating a turbulent
plane jet is described.
3.1 Context and historical review
3.1.1 Role of numerical experiments on the study of turbulence
After Orszag and Patterson (1972) demonstrated the feasibility of fully solving the Navier-Stokes
equations without the use of simplications or modelling, the new tool of Direct Numerical Simulation
(DNS) was born and has become one of the most useful in the study of turbulent ows. These computa-
tional numerical experiments have been successful in allowing easier access to all of the ow features,
unencumbered only by the researcher imagination and the prosecution of thought experiments, de-
signed to test physical theories in settings which range from the difcult too the impossible to arrange in
real, laboratory, experiments.
DNS tools are limited, however, by the available computational resources. These limitations impose
that only canonical ows of simpler geometry and moderate Re numbers are able to be simulated,
far from the high Re numbers found in ows of the natural world or of engineering interest. This is
because the wide range of scales that needs to be properly solved by the numerical simulations imply a
dependency of the number of grid points of the order of O(Re
9/4
). Nevertheless, their use has become
indispensable to confront turbulence theories and further our knowledge of it.
A insightful review of the limits, future and possibilities of DNS simulations is presented by Jimnez
(2003).
38 Direct Numerical Simulations
3.1.2 Numerical simulation of turbulent ows
As already mentioned, the large majority of ows, either natural or man made, develop at high Re and
in a turbulent state. Their simulation is thus a daunting and challenging task: as no general, analytical,
mathematical solution for the Navier-Stokes equations exists, the recourse to numerical approaches is a
must. Unfortunately, the stiff coupling of the differential equations to solve, together with the increase of
the range of scales present in these ows, in length and time, heavily limits the ow geometry complexity
and Re achievable in simulations (Moin and Mahesh (1998)). Because the previous factors directly
impact on the demanded computational and memory resources, they play a large role in the decision of
the simulation method to be employed, spanning from the most physical accurate and demanding to the
most crude and light, also according to the scientist/engineer purposes.
Fromthe several numerical simulation tools available, Direct Numerical Simulation (DNS) is the grand
queen, solving the full Navier-Stokes equations without any simplication. It allows full knowledge of the
ow characteristic variables, like velocity and pressure, as a function of time and space. However, be-
cause of the requirement to capture all the ow scales, down to the order of the Kolmogorov microscale
(O()), DNS is quite expensive in CPU and memory terms: with increasing Re, the value of the Kol-
mogorov microscale decreases, leading to an increase in grid size so a proper resolution can be kept.
This increase in number of collocation points with Re can be estimated from the ratio between the inte-
gral scale (L) and the Kolmogorov scale , i.e., , L/. Pope (2000) shows, for a HIT 3D simulation, using
a pseudo-spectral method, a scaling of N
3
Re
9/2
, being N the number of collocation points in each

direction.
The last mentioned aspect is, of course, directly connected with the amount of memory required
for the simulation to be carried through. On the other hand, the need to resolve with accuracy the
small scales present also precludes the use of large times steps in the time advancement scheme (for
a review analysis on this see Moin and Mahesh (1998)). Increasing the resolution will then decrease
the allowable time step to be used. Consequently, the wall clock time
needed to obtain the solution

increases (see, for instance, Pope (2000) for a computational cost analysis for the simple homogeneous
isotropic turbulence (HIT) case). Because of the high cost in computational resources it is very unlikely,
for the foreseeable future, that DNS will be used in an engineering context (Celani (2007); Moin and
Mahesh (1998)).
The engineer, fortunately, does not need to possess the full details of his ow problem. Instead, he
needs accurate enough values, in an average sense, for his design variables (e.g., , friction drag in
wings). In this respect he will resort to methods more amenable to the existing computational resources:
methods like Reynolds-Averaged Navier-Stokes (RANS), Large Eddy Simulation (LES) or Detached
Eddy Simulation (DES). Instead of full, direct time/space description for the ow variables, as in DNS,
these strategies are based on a statistical evolution of the ow(Moin and Mahesh (1998)). In RANS, one-
point moments like mean velocity and turbulent kinetic energy are calculated. In LES, the large energy
containing scales are directly calculated fromthe Navier-Stokes equations while the small scales solution
is obtained through some model. DES, for last, uses RANS near the wall and a LES methodology far
Running time of the simulation for the user.

3.1 Context and historical review 39
from the wall, lying somewhere between both methods.
Albeit the engineer has no interest in knowing the ow in all the detail provided by a DNS simula-
tion, he does need that the above methods - (RANS, LES, DES) - provide accurate, correct results.
His desire is that the uncertainty of the enough margin mentioned will tend to zero. One of the roles
of DNS is then to help the development, perfecting and validation of such less demanding methods,
while catering to the expansion of knowledge about the fundamental physics related with the turbulent
phenomena. In this sense, DNS is a true numerical experimental tool, going hand-in-hand with other
experimental methods, of a more classical nature, done with real ows. Of the latter, however, it can
be said that DNS has the advantage of allowing access to all the ow properties (in the measure the
numerical experimenter imagination and craft so allows), without the limitations found with experimental
apparatus. It also enables the realisation of thought experiments, useful to ascertain theoretical hypoth-
esis, impossible to translate into a classical laboratory setup (see, apropos, the comments in Jimnez
(2003)).
Being computationally intensive, the range of ows simulated using DNS has naturally expanded with
the growth of computer capabilities (some would say, nevertheless, at a despairing slow pace Celani
(2007)). The rst DNS simulations were done in the early 1970s at the National Center for Atmospheric
Research by Orszag and Patterson (1972): an HIT simulation with 32
3
collocation points and Re
of
Re
= 35, using a CDC 6600 computer. An infant by today standards, it was undoubtedly a fundamen-
tal mark, demonstrating how three-dimensional turbulent ows could be simulated with pseudo-spectral
methods. Almost 40 years later Kaneda and Ishihara (2006) reached Re
= 1200 in an HIT simula-

tion, using 4096 collocation points on the Earth Simulator, a feat some would have thought impossible
some decades ago. HIT simulations are, nevertheless, the most straightforward in the DNS realm, lack-
ing some of the numerical difculties and challenges found on non-homogeneous ows that also have
occupied the attention of the turbulence community. Namely in spatial simulations, the presence of a
non-periodic direction implies the need for a different spatial discretisation scheme. Growth in Re and
resolution for these other type of ows has not been so generous comparatively to HIT simulations (Moin
and Mahesh (1998); Dewan et al. (2005)).
3.1.3 Simulation of turbulent jets
Contrary to the number of experimental studies on plane jets
, the number of DNS and LES simula-

tions has been scarce. The rst three-dimensional spatially evolving simulation of a plane jet using LES
was performed by Dai et al. (1994). Eight years later Stanley et al. (2002) did the rst DNS simulation
of a three-dimensional spatially evolving plane jet and was, until the work here presented, the biggest
DNS of plane jet in existence. They used 19 million collocation points (390 390 130), going as farther
as 13.5 slot widths (h) with a Re = 3000, doing their computations in a Cray T3E. da Silva and Mtais
(2002a) performed plane jet simulations with 11 and 12 million points, going as far as 12.5 slot widths.
da Silva and Mtais (2002b) did the rst DNS of a round jet for Re
D
= 1.5 10
3
to Re
D
= 5 10
4
,
extending to 12.25D, i.e., inlet diameters.
For a review on experimental research on plane jet see, for instance, Dewan et al. (2005).
It has long been acknowledged that to reach the memory and CPU power needed for higher Re
and/or reach more rened resolutions to better depict all of the ow scales, use of a parallel code is
quintessential. The rst parallel simulation of an incompressible turbulent ow was done by Moin and
Kim (1982) in a ILLIAC IV machine (Hamman et al. (2007)). This was, nevertheless, a LES simulation.
The rst parallel DNS simulation of a turbulent ow was the channel ow presented in Moin and Mahesh
(1998) using a four million point grid (192 129 160) and Re = 3300. Since then, the largest channel
ow simulation is that of Hoyas and Jimenez (2006), using about 1800 million points (6144 633
4608) to attain Re
= 2003
. The parallel computational problems faced when solving the channel

ow using pseudo-spectral schemes are similar to those of simulating a jet, namely the existence of
two dimensional planes, discretised in the Fourier space, and the use of another numerical scheme,
allowing non-periodic boundary conditions, on the third direction (nite-differences, for example). Other
ows bearing the same strategy are those of stratied turbulent ows (Garg et al. (1997)).
Studying the works of Garg et al. (1997), Luchini and Quadrio (2006) and Hoyas and Jimenez (2006),
it can be seen that several strategies for domain partition and resolution of the FFTs and nite-difference
schemes are possible, but also that their efciency is dependent on the underlying parallel architecture.
Of special relevance are those considerations pertaining the use or not of parallel fast Fourier transform
(FFT) and parallel methods to solve the compact-difference scheme.
Garg et al. (1997) studied several partitioning schemes for the FFT in a distributed memory ma-
chine. Their code had, however, second-order nite differences in the third direction, which only required
near-neighbour communication. The solution presented by Luchini and Quadrio (2006) for solving the
compact-difference scheme using a version of the parallel pipelined Thomas algorithm was deemed
dependent of a shared-memory type of parallel machine or with the presence of a high-performance
communication layer as inniband or myrinet. Hoyas and Jimenez (2006) described a massive parallel
channel ow code that used the same type of domain partition adopted in the course of this work (see
chapter 4) and also implemented separated I/O nodes. From the foregoing discussion it can be seen
that the development of parallel codes, although the recourse to numerical libraries, is still machine/ar-
chitecture dependent.
3.2 Governing Equations
As stated, this work concerns the study of turbulent incompressible ows, i.e., those where /t =
0, being the uid density (mass per volume unit). Their behaviour is described by the unsteady,
incompressible Navier-Stokes equations for linear momentum eq. (2.3) and mass conservation eq.
(2.4):
Du
Dt
=
u
t
+ (u ) u =
1
p +
2
u (2.3)
Re
is the Reynolds number based on the friction velocity u
and the channel half-width. The friction velocity is dened as

u
2
=
ij
/, where
ij
is the wall shear stress.
3.3 Computational domain and boundary conditions 41
u = 0 (2.4)
Equations eq. (2.3),eq. (2.4) are valid at each point of a reference system with coordinates x =
x, y, z. The remaining variables are the velocity u, time t, pressure p(x, t) and kinematic viscosity .
The ow is further characterised by its Reynolds number Re, a non-dimensional parameter repre-
senting the balance between the inertial and viscous forces present in the ow:
Re
L
c
=
U
c
L
c
(3.1)
In eq. (3.1) U
c
is a characteristic ow velocity and L
c
a characteristic length scale. In the present work,
U
c
= U
0
is the maximum streamwise velocity and L
c
the size of the inlet aperture, H.
Simulation of turbulent ows falls globally under two approaches: temporal and spatial simulations. In
a temporal simulation, periodic conditions are applied on all boundaries, reecting a domain that follows
the ow in the streamwise direction (e.g., Lagrangian approach). Spatial simulations, by opposition, have
a xed computational domain, needing inow and outow boundary conditions (Eulerian approach).
a) b)
Figure 3.1: a) Temporal plane jet (pressure iso-surfaces). The exiting structures entering the domain again through
the inlet. b) Spatial plane jet (vorticity iso-surfaces coloured by velocity magnitude). Jet develops in space from the
inlet, boundary conditions at the exit allow uid to escape the computational domain.
The periodic boundary conditions of temporal simulations impose restrictions on the size of the ow
domain, generally under predicting the shear layer growth due to the pressure feedback mechanism
that causes small downstream perturbations which affect the upstream ow. Spatial simulations tend to
be more realistic but are, simultaneously, more computational demanding: the full ow domain must be
accounted for, from jet inlet to transition to self-similar zone. In addition, special care must be taken to
properly address the inow and outow boundary conditions. The former must properly model realistic
initial ow conditions, the latter must allow correct exit of the ow. The high computational cost associ-
ated with spatial simulations have so far precluded the existence of very large simulations. Addressing
that situation is one of the goals of the present work.
3.3 Computational domain and boundary conditions
The computational domain for the spatial simulation consists of a box with L
x
,L
y
and L
z
dimensions
in each of the x (streamwise), y (normal) and z (spanwise) directions (see g. (3.2)). The y and
z directions have periodic boundary conditions and the streamwise direction has inow and outow
boundaries, as shown in the aforementioned gure.
Figure 3.2: Computational domain.
3.4 Spatial discretisation
3.4.1 Pseudo-spectral scheme
Due to their high-level of accuracy, pseudo-spectral schemes are in a prominent position for answer-
ing DNS needs (Canuto et al. (1987)). Additionally, for the same resolution, they have a computational
cost proportional to nlog
2
n, where n is the number of discretisation points used. This cost is quite low
comparing with other numerical methods (Ferziger and Peri ` c (1996)). They are limited, however, to the
treatment of periodic ows.
Given a periodic ow variable, (x, y, z, t), along the normal (y) and spanwise (z) directions, an
inverse 2D discrete Fourier transform can be used to expand it through eq. (3.2),
(x, y, z, t) =
n
y
2
1
j=
n
y
2
n
z
2
1
k=
n
z
2
(x, k
y
, k
z
, t)e
(k
y
y+k
z
z)
(3.2)
where j, k are the indexes of the points along each direction y, z and (k
y
, k
z
) are the Fourier wave
numbers dened by,
k
y
=
2
L
y
j
k
z
=
2
L
z
k. (3.3)
where i =
1 is the imaginary unit, L

y
and L
z
the spatial dimensions along the y and z directions and
n
y
and n
z
the number of discrete collocation points along those same directions, respectively.
Each Fourier coefcient

(x, y, z, t), can then be obtained using the (direct) 2D discrete Fourier
transform,
(x, k
y
, k
z
, t) =
1
n
y
n
z
n
y
1
j=0
n
z
1
k=0
(x, y, z, t)e
(k
y
y+k
z
z)
(3.4)
3.4 Spatial discretisation 43
The derivatives in the physical space result from the simple multiplication in the Fourier space, like
y
= k
y

z
= k
z

(3.5)
3.4.2 Compact nite differences discretisation
Because pseudo-spectral formulation imposes periodic conditions on the computational domain
boundaries, another discretisation scheme is needed for the streamwise derivatives. To safeguard the
requirement of high resolution, a compact formulation is used (Lele (1992)). This formulation consists in
expressing the derivative at a given point as a function of its neighbour points derivatives. It is an implicit
scheme, resulting in the need to solve a linear system of equations to extract any spatial derivative.
These schemes provide the required high order accuracy, possessing to some measure good spectral
characteristics.
Considering an uniform grid with spatial coordinates x
i
= (i 1)x, with x = Const. (i [1, n
x
],
x
i
[0, L
x
]), and using Taylor expansions of the function f
i
= f(x
i
) = f(x) with rst derivative f
i
=
f
(x
i
) =
df
dx
(x
i
) and second derivative f
i
= f
(x
i
) =
d
2
f
dx
2
(x
i
), the general formula can be arrived at,
p
j=p
j
f
i+j
=
q
k=q
a
k
f
i+k
+ O(x
n
) (3.6)
with p, q N. The order of the approximation n depends on the restrictions imposed upon the numbers
j
and a
k
. Particular cases are:
j ,= 0 ,
j
= 0 > explicit scheme.
j ,= 0 , with
j
= 0 > Hermitian scheme (Compact).
(j, k) ,
j
=
j
and a
k
= a
k
> centered scheme.
(j, k) with
j
,=
j
and/or a
k
,= a
k
> discentered scheme.
For instance, using p = 2 and q = 3 in eq. (3.6), eq. (3.7) is obtained.
f
i2
+ f
i1
+ f
i
+ f
i+1
+ f
i+2
= a
f
i+1
f
i1
2x
+ b
f
i+2
f
i2
4x
+ c
f
i+3
f
i3
6x
+ O(x
n
) (3.7)
to which the following restrictions on (, , a, b, c) apply:
a + b + c = 1 + 2 + 2 if n 2.
a + 2
2
b + 3
2
c = 2
3!
2!
( + 2
2
) if n 4.
a + 2
4
b + 3
4
c = 2
5!
4!
( + 2
4
) if n 6.
...
The calculation of each streamwise derivative depends on each particular point location. All interior
points are calculated with a central compact formulation,
f
i1
+ f
i
+ f
i+1
= a
f
i+1
f
i1
x
+ b
f
i+2
f
i2
x
(3.8)
For those points between i = 3, ..., n
x
2, 6th order accuracy was attained by setting = 1/3 and
enforcing the following restrictions on a and b,
a = ( + 2)/3 = 7/9 , b = (4 1)/12 = 1/36. (3.9)
For those points near the two boundaries, along the streamwise direction (i = 2 and i = n
x1
), a
classical 4th order Pad scheme was used, corresponding to setting = 1/4 and,
a = ( + 2)/3 = 3/4 , b = (4 1)/12 = 0. (3.10)
Finally, decentered, 3rd order implicit schemes were used for the points lying at the boundaries (i = 1
and i = n
x
),
f
1
+ 2f
2
=
1
2x
(5f
1
+ 4f
2
+ f
3
) (3.11)
f
n
x
+ 2f
n
x
1
=
1
2x
(5f
n
x
4f
n
x
1
f
n
x
2
) (3.12)
In practice, the calculation of a single derivative at given point x
p
= (x
p
, y
p
, z
p
) consists in solving a
matrix equation involving all the x derivatives along the (x, y
p
, z
p
) direction. The system is thus written
as,
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
1
1
2
1
2
1
.
.
.
.
.
.
.
.
.
1
.
.
.
.
.
.
.
.
.
1
n1
1
n1
n
1
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
f
1
f
2
f
3
.
.
.
f
i
.
.
.
f
n
x
2
f
n
x
1
f
n
x
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
=
1
x
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
a
1
b
1
c
1
a
2
0 a
2
b a 0 a b
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
b a 0 a b
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
b a 0 a b
a
n1
0 a
n1
c
n
b
n
a
n
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
f
1
f
2
f
3
.
.
.
f
i
.
.
.
f
n
x
2
f
n
x
1
f
n
x
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
(3.13)
with the following coefcients, as discussed above,
1
=
n
= 2 , a
1
= a
n
=
5
2
, b
1
= b
n
= 2 , c
1
= c
n
=
1
2
3.4 Spatial discretisation 45
2
=
n1
=
1
4
, a
2
= a
n1
=
3
4
=
1
3
, a =
7
9
, b =
1
36
The second streamwise derivatives has an equation analogous to eq. (3.14) where f
i
> f
i
is now,
f
i2
+f
i1
+f
i
+f
i+1
+f
i+2
= a
f
i+1
2f
i
+ f
i1
x
2
+b
f
i+2
2f
i
+ f
i2
4x
2
+c
f
i+3
2f
i
+ f
i3
9x
2
+O(x
n
)
(3.14)
and the restrictions on (, , a, b, c) become:
a + b + c = 1 + 2 + 2 if n 2.
a + 2
2
b + 3
2
c = 2
4!
2!
( + 2
2
) if n 4.
a + 2
4
b + 3
4
c = 2
6!
4!
( + 2
4
) if n 6.
...
Like for the rst derivatives, all interior points were calculated with a central compact formulation
which, for the second derivative is
f
i1
+ f
i
+ f
i+1
= a
f
i+1
2f
i
+ f
i1
x
2
+ b
f
i+2
2f
i
+ f
i2
x
2
(3.15)
and it is corresponds to a 6th order accuracy for the points laying between i = 3, ..., n
x
2. This is
equivalent to have = 2/11 and,
a = 4(1 )/3 = 12/11 , b = (4 1)/12 = 3/132. (3.16)
Like before, points at i = 2 and i = n
x1
are treated with a classical 4th order Pad scheme. For the
second derivative this is recovered with = 1/10 and,
a = 4(1 )/3 = 12/10 , b = (4 1)/12 = 1/20. (3.17)
A decentered, 3rd order implicit scheme was used for the boundary points at i = 1 and i = n
x
,
f
1
+ 11f
2
=
1
2x
2
(13f
1
27f
2
+ 15f
3
f
4
) (3.18)
f
n
x
+ 11f
n
x
1
=
1
2x
2
(13f
n
x
27f
n
x
1
+ 15f
n
x
2
f
n
x
3
) (3.19)
The nal system of equations results in,
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
1
1
2
1
2
1
.
.
.
.
.
.
.
.
.
1
.
.
.
.
.
.
.
.
.
1
n1
1
n1
n
1
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
f
1
f
2
f
3
.
.
.
f
i
.
.
.
f
n
x
2
f
n
x
1
f
n
x
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
=
1
x
2
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
a
1
b
1
c
1
d
1
a
2
2a
2
a
2
b a 2(a + b) a b
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
b a 2(a + b) a b
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
b a 2(a + b) a b
a
n1
2a
n1
a
n1
d
n
c
n
b
n
a
n
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
f
1
f
2
f
3
.
.
.
f
i
.
.
.
f
n
x
2
f
n
x
1
f
n
x
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
(3.20)
with the coefcients,
1
=
n
= 11 , a
1
= a
n
= 13 , b
1
= b
n
= 27 , c
1
= c
n
= 15 , d
1
= d
n
= 1
2
=
n1
=
1
10
, a
2
= a
n1
=
12
10
=
2
11
, a =
12
11
, b =
3
44
Eq. (3.13) and eq. (3.20) are more simply expressed by
A
1
f
=
1
x
B
1
f (3.21)
A
2
f
=
1
x
2
B
2
f (3.22)
where A
1
, A
2
, B
1
and B
2
are (n
x
n
x
) matrices and f , f
and f
are (n
x
1) vectors.
Having dened the matrices A
1
, A
2
, B
1
and B
2
, the computation procedure to get both vectors f
and f
from vector f , can be thus described:

1. Inversion of A
1
and A
2
matrices,
A
1
1
, A
2
1
(3.23)
2. Multiplication of matrices B
1
and B
2
by the vector
1
x
f ,
_
1
x
B
1
f
_
,
_
1
x
2
B
2
f
_
(3.24)
3. Multiplication of matrices A
1
1
and A
2
1
by the vectors
1
x
B
1
f and
1
x
2
B
2
f ,
A
1
1
_
1
x
B
1
f
_
, A
2
1
_
1
x
2
B
2
f
_
. (3.25)
The above strategy for calculating f
and f
is much cheaper than computing the inversion of the full

M
1
= A
1
1
B
1
and M
2
= A
1
2
B
2
matrices. In fact, matrices M
1
and M
2
are full and their inversion (in
order to get f
and f
) has a quite high cost - proportional to n

2
x
. Using the above procedure only the
inversion of matrices A
1
and A
2
is needed, resulting in a computational cost proportional to n
x
.
3.5 Temporal discretisation 47
3.5 Temporal discretisation
The simulation results from solving equations eq. (2.3) and eq. (2.4) in space and time. The time
evolution is obtained using an explicit third order Runge-Kutta scheme. Equations eq. (2.3),eq. (2.4)
can be re-written as:
u
t
= N(u) + L(u) (3.26)
u = 0 (3.27)
where N(u) is the convective term
N(u) = u (3.28)
and L(u) represents the viscous term
L(u) =
2
u (3.29)
(x, t) the vorticity vector
= u (2.1)
and (x, t) the modied pressure given by
=
p
+
u
2
2
(3.30)
The 3 step, 3
rd
order Runge-Kutta scheme, calculates each sub-step velocity p from the last two time
sub-steps p 1 and p 2 using
u
p
u
p1
t
=
p
_
N(u
p1
) + L(u
p1
)
+
p
_
N(u
p2
) + L(u
p2
)
+
P
(3.31)

p
= 0 (3.32)
The coefcients in eq. (3.31) are, for a 3
r
d order accuracy (Willianson, 1980) with 3 time sub-steps
p = 1, 2, 3 are
1
= 8/15
1
= 0
2
= 5/12
2
= 17/60
3
= 3/4
3
= 5/12
(3.33)
To solve eq. (3.31), the pressure eld must be known at each time sub-step p. Equation eq. (3.32),
on the other hand, doesnt give the pressure eld in an explicit way but rather enforces a constraint
inherent to all incompressible ow elds. The role of the pressure is to enforce this constraint, thus
leading to a strong coupling between eq. (3.31) and eq. (3.32) and the need for solving both equations
simultaneously.
3.5.1 Pressure-velocity coupling
The fractional step method (Kim and Moin (1985),Le and Moin (1991)) is used here to solve the
simultaneous calculation of the velocity and pressure eld. This is necessary to insure incompressibility
of the velocity eld at each time sub-step and it implies that a Poisson equation must be solved at each
sub-step of the Runge-Kutta scheme.
Using u
p
for the sub-step velocity eld, a new velocity eld u
can be obtained as
u
u
p1
t
=
p
_
N(u
p1
) + L(u
p1
)
+
p
_
N(u
p2
) + L(u
p2
)
+ (3.34)
The pressure eld at the end of sub-step p has then to obey

p
t
=
2
p
(3.35)
which leads to
p
=
u
t
(3.36)
The algorithm can then be resumed as
1. Obtain u
just through the advancement of the convective and viscous terms from eq. (3.34). Note
that the eld u
does not satisfy incompressibility ( u
= 0);
2. Solve eq. (3.35) to get the pressure eld of sub-step p;
3. Obtain the correct velocity from
u
p
u
t
=
p
(3.37)
3.5.2 Spatial discretisation
Except for multiplications, all the calculations are done in the Fourier space. Equation eq. (3.31) is
thus rewritten as
u
p
u
p1
t
=
p
_
N
p1
+ (
d
2
dx
k
2
) u
p1
_
+
p
_
N
p2
+ (
d
2
dx
k
2
) u
p2
_
(3.38)
where
N
p
=
_
_
u
p
y
p
z
u
p
z
p
y
u
p
z
p
x
u
p
x
p
z
u
p
x
p
y
u
p
y
p
x
_
_
(3.39)

p
x
= ik
y
u
p
z
ik
z
u
p
y

p
y
= ik
z
u
p
x
d u
p
z
dx

p
z
=
d u
p
y
dx
ik
y
u
p
x
(3.40)
and
k
2
= k
2
y
+ k
2
z
(3.41)
After solving eq. (3.38), the correction pressure eld
p
is obtained solving the Poisson problem (also in
the spectral space)
_
d
2
dx
2
k
2
_

p
=
1
t
_
d u
x
p
dx
+ ik
y
u
y
p
+ ik
z
u
z
p
_
(3.42)
3.6 Boundary conditions 49
The velocity eld u
, obtained in eq. (3.38), is then corrected by

u
x
= u
p
x
t
_
d
p
dx
_
u
y
= u
p
y
t (ik
y

p
)
u
z
= u
p
z
t (ik
z

p
)
(3.43)
The velocity eld obtained from eq. (3.43) then satises the incompressible condition
d u
p
x
dx
+ ik
y
u
p
y
+ ik
z
u
p
z
= 0 (3.44)
3.5.3 Stability condition
The time step used in the numerical resolution of a system of partial differential equations is bounded
by stability restrictions (see, for instance, Ferziger and Peri ` c (1996)). The key parameter governing the
numerical method stability, in convection dominated problems (such as those of interest here), is the
Courant number expressed by eq. (3.45)
C
fl
=
[u[t
[x[
(3.45)
The limiting Courant number for a particular problem is many times established experimentally, and, in
the multidimensional case, an assessment of eq. (3.45) done for each of the ow directions. For the
spatial plane jet case described, a limit value of C
fl
< 0.7 was found. At each iteration the maximum
velocity value of each component is found and if the former condition is violated, t re-adjusted using
eq. (3.46).
t = C
fl
min
_
x
[u
x
[
max
,
y
[u
y
[
max
,
z
[u
z
[
max
_
(3.46)
In eq. (3.46) [u
x
[
max
, [u
y
[
max
, [u
z
[
max
is the maximum velocity in each direction found in the oweld.
The factor in the y and z direction is due to the spectral nature of the spatial discretisation in those
directions (see Canuto et al. (1987)).
3.6 Boundary conditions
Like in all numerical codes, the proper implementation of boundary conditions is crucial for their
success. The following paragraphs detail the boundary conditions implementations for the spatial plane
jet.
3.6.1 Inlet
A velocity prole is imposed at the inlet at each timestep, having a general form of
u(x
0
, t) = U
med
(x
0
) +U
noise
(3.47)
x
0
= (x = 0, y, z) is the coordinate at the inlet plane, U
med
the mean inlet prole and U
noise
a superim-
posed random numerical noise. For the planar jet, the average y, z components of mean inlet velocity
prole are zero and, in the x direction, a tangent hyperbolic prole (see Ribault et al. (1999); Stanley
et al. (2002)) expressed by eq. (3.48) is used.
U
med
=
U
1
+ U
2
2
+
U
1
U
2
2
tanh
_
h
2
0
y
H
_
(3.48)
In eq. (3.48) is the momentum thickness of the initial shear layer and H the inlet slot width. U
1
is the
centreline velocity and U
2
a co-ow velocity. This co-ow must be added to allow the growth of the jet,
otherwise hampered by the lack of natural entrainment from the lateral boundaries. U
2
needs to be only
high enough to allow for this growth, otherwise it would affect the instability characteristics of the initial
mean velocity prole.
Numerical noise, U
noise
, is also superposed to the mean prole in eq. (3.48), at each timestep t.
This noise prole has the form of
U
noise
= ampU
base
u
(3.49)
where amp is an amplication factor generally ranging from [0, 0.15] and u
= (u
, v
, w
) comes from a
random number generator designed to conform to the energy spectrum
E(k) k
s
exp
_
s
2
(k/k
0
)
2
_
(3.50)
where s is the slope of the noise spectrum in the large scales region and, k
0
the wave number location
of maximum every. Both parameters are chosen so to provide the best realistic conditions. The choice is
dictated so to allow a natural transition of jet, opting by a largely white noise prole, with even importance
given to all wave numbers. The choice of the modes to be amplied by the transition process is then left
to the jet itself. The numerical noise is convolved through the U
base
, setting the noise location mainly in
the shear layer gradients region.
3.6.2 Lateral boundaries
The lateral boundaries are of periodic nature, a direct consequence of the choice of the spectral
discretisation in the y and z directions of the ow:
u(x, y, z, t) = u(x, y + L
y
, z, t)
p(x, y, z, t) = p(x, y + L
y
, z, t)
u(x, y, z, t) = u(x, y, z + L
z
, t)
p(x, y, z, t) = p(x, y, z + L
z
, t)
(3.51)
As can be seen by eq. (3.51), no enforcement of zero normal velocity is imposed although the mean
normal velocity is zero. Also, as mentioned earlier, these boundary conditions allow for the natural
jet growth. To suppress this problem the lateral boundaries must be placed sufciently far way and a
small co-ow added to the inlet prole (see Stanley and Sarkar (2000a)). If this is properly done, the
jet dynamics will develop naturally. The distance of the lateral boundaries is also dictated by the need
to avoid spurious reections which would upset the ow. To evaluate if this distance is large enough,
several tests must be done and the results compared to those existing on the literature (CITES). It must
also be noted that the aforementioned problems only affect the normal direction, the spanwise direction
is naturally periodic.
Figure 3.3: Lateral boundary condition. Side view of turbulent plane jet illustrated by vorticity iso-surfaces and
indicating streamwise position of Y 0Z slice. On the right two views of velocity vectors of said slice: a) shows
normalised velocity vectors by their own magnitude to display velocity arrangement in space. b) shows velocity
vectors proportional to their magnitude, scaled for better viewing. It can be seen from b) that the co-ow velocity is
minimal.
3.6.3 Outow
A correct implementation of the outow condition is the crux of achieving a successful simulation:
it must ensure that the coherent structures are not be affected by the outow condition when exiting
the domain. Non-reective boundary conditions where all the terms of the Navier-Stokes equations -
convective and viscous - are explicitly advanced where thus used Orlanski (1976). Because pressure
and velocity are strongly coupled they are also simultaneously advanced, ensuring that their values are
compatible. The u
x
momentum equation is modied and the longitudinal pressure gradient,
p
x
, and the
longitudinal dissipation term, Re
1
2
u
x
2
, included in a new variable C
u
(y, z). This variable also accounts
for the longitudinal convective terms C
u
(y, z)
u
x
. The equation nal shape is thus
u
x
t
= C
u
x
(y, z)
u
x
x
u
y
u
x
y
u
z
u
x
z
+ Re
1
_

2
y
2
+

2
z
2
_
u
x
(3.52)
The value of u
x
is obtained at each Runge-Kutta sub-timestep p. First eq. (3.52) is used to calculate
C
u
x
(y, z) at i = n
x
1:
u
p1
x
u
p2
x
t
= C
u
x
(y, z)
u
x
x
u
y
u
x
y
u
z
u
x
z
+ Re
1
_

2
y
2
+

2
z
2
_
u
x
(3.53)
The longitudinal derivative is calculated with a 1
st
order backward differencing scheme while the normal
and spanwise directions with a the pseudo-spectral scheme. Then the same equation is used with the
Runge-Kutta time scheme to obtain the streamwise velocity at the outlet plane. After this the Poisson
equation is solved and, nally, u
y,z
corrected at i = n
z
with
u
y
(y, z) = u
y
(y, z)
p
y
u
z
(y, z) = u
z
(y, z)
p
z
(3.54)
where u
y,z
are the velocity components obtained from the Navier-Stokes equations without the pressure
gradient term. It has been proven that this scheme guarantees a smooth exit of the coherent structures,
with negligible distortion from the presence outow boundary condition (see g. (3.4)).
Figure 3.4: Coherent structures outow. Several snapshots of the progress of coherent structures until they exit the
computational box. Pressure iso-surfaces, with Pr = 0.05 was used for illustration purposes.
3.6.4 Poisson equation
The boundary conditions applied to solve the Poisson equation are also, for the spanwise and normal
directions, periodic, due to the pseudo-spectral scheme. For the inlet and outlet, Neumann type condi-
tions were imposed. This implies the need of knowing the pressure gradient outside of the computational
domain (for x [
x
, 0] and x [L
x
, L
x
+
x
] ):
p
x
[
x
,0]
=
p
x
i=1
(3.55)
p
x
[L
x
,L
x
+
x
]
=
p
x
i=n
x
(3.56)
These longitudinal pressure gradients are obtained from velocity elds at the inlet and outlet, respec-
tively. At the inlet, the imposed velocity u
x
is used in conjunction with the velocity from the Navier-Stokes
equations, without the pressure gradient term, u
x
:
t
p
x
i=1
= u
i=1
u
i=1
(3.57)
and the outlet, eq. (3.58) uses the advanced velocity u
x
from eq. (3.53)
t
p
x
i=n
x
= u
i=n
x
u
i=n
x
(3.58)
For discretisation purposes, eq. (3.55), eq. (3.58) corresponds to
f
0
= f
1
(3.59)
f
n
x+1
= f
n
x
(3.60)
where 0 and n
x
+ 1 are ctitious nodes placed outside the computational domain. Consequently the set
of expressions in eq. (3.61) is obtained, which allows for the inversion of the Poisson equation.
f
1
=
8
7
f
2
1
7
f
3
6
7
xf
1
f
0
= f
1
xf
1
f
n
x
=
8
7
f
n
x
1
1
7
f
n
x
2
6
7
xf
n
x
f
n
x
+1
= f
n
x
xf
n
x
(3.61)
CHAPTER 4
Parallel plane jet
This chapter describes the details pertinent to the parallel implementation of the algorithm described in
the previous chapter. Discussion is also provided in particular points about the available options and
choices made.
It is unquestionable that to answer the needs of achieving simulations with higher resolutions, higher
Re, higher complexity computational simulation codes must be written for parallel architectures. This
was true in the 70s and is more so now, when even in the consumer market the road of increasing cores
has replaced that of increasing CPU frequency. Coupled with the numerical challenge, DNS simulations
are also a problem of software engineering, where the underlying machine architecture, characteristics
and limitations must be taken into account to achieve effectual programs.
While porting the turbulent plane jet DNS code described in chapter 3 to a parallel environment
several choices were made that impacted on the code design. These decisions were guided by:
target parallel machines where the code should run;
performance;
modular approach to ease the addition of new models/uses (implementation of LES, different
boundary conditions, passive scalar.
The following chapter describes how these objectives shaped the decisions taken and provides a
description of the parallel code architecture and its implementation.
4.1 Parallel architecture
It has become clear that in the last decade the vast majority of High Performance Computing (HPC)
machines are of the cluster typology. In November 2010 TOP 500 (2010), a website compiling the 500
56 Parallel plane jet
most powerful computers in the world, showed that 82.8% of those where of the cluster type.
Almost all clusters are of the Beowulf type (or derivatives): they use commercial off-the-shelf (COTS)
machines as computing nodes, connected by a dedicated network fabric (ethernet, inniband, myrinet,
etc. . . ) and running an open-source software stack (TOP 500 (2010) reported, for Nov. 2010, that 91.8%
of its machines were operating some sort of Linux avour).
Albeit the cluster denomination covers the majority of parallel machines in use today, it is still a loose
term, leaving outside several details about the actual layout and interconnection of components. The
availability or not of a parallel lesystem with optimised parallel I/O, the type of dedicated interconnect
used, the existence of dedicated I/O nodes, the number of core count per node, if these are of the NUMA
type or not, etc. . . . These factors imply several code design decisions when ultimate performance is the
goal. Fortunately, several libraries exist that create an abstraction layer that allow a certain degree of
decoupling between the algorithm and machine. Still, the particular architecture details have always to
be accounted for.
Figure 4.1: Example of cluster layout. Exterior access is provided through the master node. In larger clusters
independent login nodes are established to provide extra-security. Storage is available through a parallel lesystem
(Gluster, GPFS, Lustre, etc. . . or NFS). Computing nodes are also connected with a dedicated high performance
network for data exchange (generally through MPI).
The present implementation is aimed towards the most probable cluster type machines to be have
access to, of which a typical, general layout is given in g. (4.1).
One of the most signicant factors introduced by parallel computing is to give emphasis to inter-
process communication dependencies. Striving for locality, minimising data exchanges between pro-
cesses, is then one of the general thumb rules of parallel programming as will be seen further in the
text.
4.2 Parallel DNS 57
4.2 Parallel DNS
Inasmuch the code must be run in different realisations of the cluster type of architecture and, in
many of those there will be no possibility of tweaking or installing dedicated software, a decision was
made of using generally available library tools for portability and convenience. The rst consequence
was the option to design the code around the message passing paradigm as described by the MPI
standard Mes (1997). This ensures that the code can be effortlessly ran in the maximum number of
cluster types.
Although several MPI implementations exist, from open-source to highly optimised proprietary ven-
dor libraries, some general deciencies are found across them: for instance, difculty of having real
ability to overlap computation and communication or small efciency in communicating small messages
efciently. This problems were also considered in the design of the parallel planar jet DNS version.
As mentioned in the previous chapter, pseudo-spectral (and fully spectral) parallel DNS codes are,
in general, highly communication intensive. This hampers their scalability and parallel efciency by
Amdahls law. Albeit not many references were found to the parallel plane jet DNS simulation codes, the
algorithm is of the same nature than those found in pseudo-spectral channel ows simulations, where
more extensive references exist.
Also from the introductory section of the previous chapter, it is clear that an efcient mix between
operations related with the compact differences discretisation scheme, and those pertaining the spectral
operations, must be stricken. That is the subject matter of the next sections.
4.3 Data dependency analysis
To ascertain an adequate parallel strategy, the data dependencies implied in the numerical scheme
must be understood. They impact directly on the solutions to be devised to achieve a proper imple-
mentation. This implies close attention to the target parallel machine architecture and goals of the code
(shall it only run in a specic dedicated architecture or should it be more exible?). As the target is,
generally speaking, distributed memory machines, some sort of data domain division is implied. In the
present case a mixed type of domain division was decided upon, optimising the different requirements
of the two types of discretisation used, spectral and nite-differences. To better gauge the implications,
a quick review of the numerical scheme previously explained follows.
From the previous section the numerical scheme can be summed up, for each Runge-Kutta sub-step,
in a three stage progression:
i. solve the the right hand side:
a calculate the nonlinear terms
b calculate the viscous terms and obtain velocity
ii. obtain the correction for the velocity from the pressure (through Poisson)
4.3.1 Right hand side:
u
t
= N(u) + L(u)
4.3.1.1 Non linear term
The right hand side involves solving the convective, nonlinear term N(u) (see eq. (3.28)) and the
viscous term L(u) (eq eq. (3.29)):
u
t
= N(u) + L(u) (4.1)
The rst thing is to calculate the nonlinear term in the spectral domain. This requires the computation
of several rst order derivatives in all directions in order to obtain the (eq. (3.40), repeated below for
convenience).

p
x
= ik
y
u
p
z
ik
z
u
p
y

p
y
= ik
z
u
p
x
d u
p
z
dx

p
z
=
d u
p
y
dx
ik
y
u
p
x
(3.40)
All derivatives in the streamwise direction (in this case u
p
y
/x, u
p
z
/x) use the compact differences
scheme. This signies the resolution of n
y
n
z
independent pentadiagonal matrices, or, speaking data
wise, n
y
n
z
independent lines of size n
x
. The derivatives in the normal and spanwise direction are
calculated in the spectral space, with the data independent by ZoY planes, meaning n
x
independent
planes of size n
y
n
z
. This tight coupling pattern of two independent local data topologies spans the rest
of the sub-timestep algorithm: compact differences operations take place along independent lines in the
streamwise direction, and spectral operations in independent planes dened by in the normal/spanwise
directions.
The next step sends and v to the physical space to ascertain max [abs(u
x
), abs(u
y
), abs(u
z
)]
(needed to enforce the CFL stability condition (see section 3.5.3)) and calculate N(u) = u (3.39,
repeated below):
N
p
=
_
_
u
p
y
p
z
u
p
z
p
y
u
p
z
p
x
u
p
x
p
z
u
p
x
p
y
u
p
y
p
x
_
_
(3.39)
While in the physical space several monitoring quantities related to the ow energy are also calculated
(e.g., ,

n
i=1

j
j
and

n
i=1
u
j
u
j
). After this N(u) is returned to the spectral space.
4.3.1.2 Viscous term and velocity advancement
To calculate the viscous term (eq. (3.29), L(u) =
2
u), the second-order derivatives must be
obtained. Once again, the spanwise derivatives
2
/x
2
come from the compact differences scheme
and the normal and spanwise derivatives from the spectral discretisation. In the end, the following right
hand side is obtained:
N( u
p
) +
_

2
x
2
k
2
_
u
p
(4.2)
Note that, during the last step, except for the calculation of the second-order derivatives in the stream-
wise direction all operations are data independent, i.e., only the values of the variables at each point are
needed because those in the spectral space are already known.
4.4 Parallel domain decomposition 59
At the end of this step, the velocity eld is advanced inside the Runge-Kutta sub-step, but still needs
to be corrected to respect the incompressibility condition.
4.3.2 Poisson equation corrective term u
This last step enforces the incompressibility condition fromthe Poisson equation in the spectral space
(eq. (3.42), repeated below):
_
d
2
dx
2
k
2
_

p
=
1
t
_
d u
x
p
dx
+ ik
y
u
y
p
+ ik
z
u
z
p
_
(3.42)
The resolution of the Poisson equation also involves the calculation of streamwise partial deriva-
tives. Inasmuch as no space changing is involved (spectral to physical or vice-versa), the only data
dependencies are related to the presence of the compact-difference discretisation for the streamwise
derivatives.
4.3.3 Boundary conditions
The inlet boundary condition is calculated at the beginning of each global timestep. The outlet
condition, on the other hand, is applied at each Runge-Kutta sub-timestep like the Poisson boundary
condition referred in sec.3.6. Because of the presence, once more, of the streamwise derivatives, a data
dependency arises in that direction.
4.4 Parallel domain decomposition
It is clear from the previous analysis that two different data dependencies can be identied: one
associated with the compact scheme (points in the streamwise direction) and other with the FFT opera-
tions, involving the normal/spanwise planes. This is equal to other pseudo-spectral DNS codes of similar
nature and the literature has focused, in that regard, on the case of parallel channel-ow simulations. A
general overview of the various alternative implementation strategies can be summed:
4.4.1 Streamwise slabs
Figure 4.2: Slab division (streamwise planes).
One possible single domain division strategy is the use of slabs in the streamwise direction (g.
(4.2)). This keeps all the data needed for each solution of the compact derivatives and of the Poisson
equation in the same node. Solution of the FFTs related terms is obtained with a parallel algorithm.
This option implies the use of one or two transpose operations to calculate the 2D FFT (depending if
the same index ordering is desired at the end of the inverse/direct operation). The parallel transpose is
one of the most expensive operations, communication wise, because it is of the All_to_All type (i.e., all
nodes need to exchange data with all other nodes).
4.4.2 Normal slices
Figure 4.3: Y 0Z slices division (normal/spanwise planes).
Another single domain division strategy can be pursued using slices that keep the Y oZ planes intact
(g. (4.3)). This ensures that the FFT operations are kept in the same node. Resolution of the compact
scheme discretisation and Poisson problem would be done using a parallel matrix solver. Because of
precision requirements, a direct-solver is desired. The parallel pipelined Thomas algorithm (PPT) sat-
ises this requirement, when deployed in machines with high-performance interconnects, and assisted
by a communication strategy capable of dealing efciently with a big number of small messages (see
Luchini and Quadrio (2006)). Unfortunately, several issues are raised with this option in the case at
hand. PPT, to be efcient, requires a large amount of small messages to be sent and received between
the several calculating nodes. The MPI libraries in existence, on the other hand, achieve the needed
level of efciency for this requirement, being optimised for large packets of data. Luchini and Quadrio
(2006) solved this problem developing a special interconnect topology, bypassing the need for a switch,
and directly communicating through sockets using the I/O functions of C library. The use of the PPT
was then discarded due to the requirement of using MPI to ensure code portability, i.e., to be able to
run in cluster facilities managed externally and with signicant different types of interconnect hardware:
from TCP/DIP based to specic high performance interconnects like Inniband and Myrinet. Developing
a similar strategy would imply programming a dedicated communication layer to accommodate different
types of interconnects, and accept the risk of the associated communication topology of the cluster in
question preclude any possible advantages of such work.
4.4.3 Mixed domain decomposition
The opted for implementation relies thus in a mixed decomposition (g. (4.4)), where the slab and
slice conguration co-exist simultaneously during the simulation. This ensures that all calculations are
4.4 Parallel domain decomposition 61
Figure 4.4: Slab and slices mixed decomposition. Data exchange done through MPI_AlltoAll function call
local: FFT in the slice partition and the compact discretised terms in the slab partition. Like in the
slab only strategy, the data needs to be transposed from one conguration to the other, through an
AlltoAll operation (in this case using MPI_Alltoall). The price for this operation is the same as doing the
transpose for the parallel FFT, as shown by Pelz (1992) with the advantage of keeping calculations local.
One possible drawback of this conguration is the doubling in needed memory. This can be avoided
however, by allocating only the memory space needed by one of the partitions, and re-using it through
a reshape operation for the other partition. For this to function properly, both congurations must be
isolated by a buffer during the data exchanging phase. With the recent Fortran 2003 standard, this
option is facilitated by the ISO_C_BINDING intrinsic module, and its c_f_pointer and c_loc function. An
example is given below:
1 use i nt r i nsi c : ISO_C_BINDING
2 i nt eger : : ni , nj , nk , np
3 real , al l ocat abl e , dimension ( : ) : : ux
4 real , dimension ( : , : , : ) , poi nt er : : pt _ux_sl ab , pt _ux_sl i ce
5
6 ! np = number of pr ocessor s
7 ! ni , nj , nk = t o t a l di mensi ons i n x , y , z di r ec t i ons
8 ! ux x component of t he v el oc i t y
9
10 ! al l oc at e t he memory space
11 al l ocat e ( ux ( ni
*
nj
*
nk )
12 ! cr eat e poi nt er f or ux i n t he s l i c e t ype p a r t i t i o n
13 cal l c_f _poi nt er ( c_l oc ( ux ) , pt _ux_sl i ce , [ nj , nk , ni / nproc ] )
14 ! cr eat e poi nt er f or ux i n t he sl ab t ype p a r t i t i o n
15 cal l c_f _poi nt er ( c_l oc ( ux ) , pt _ux_sl ab , [ ni , nj / nproc , nk ] )
Listing 4.1: Fortran 2003 use of c_loc and c_f_pointer
With the above example code, both pt_ux_slab and pt_ux_slice will use the same memory space,
although with different topologies. Examining the code above, it is easy to notice the change in index
ordering from i, j, k for the slabs to j, k, i for the slices. This re-ordering ensures memory continuity
and better cache performance when doing the FFT calculations and solving the compact discretised
derivatives and the Poisson equation. This factor accounts for almost a two-fold increase of performance
versus using the same ordering in all partitions i, j, k.
Another point worth of note is the choice of directions for the domain division. If the division along
the streamwise axis is a straightforward matter, that along the normal Y axis needs a comment:
for the plane jet the relation n
j
>> n
k
generally holds true. Knowing that the domain division must be
carried away for n
proc
processors, using the normal direction allows to maximise the number of possible
central processing units (CPUs) to run (the current implementation does not allow to use a number
of processors greater than number of discretisation points on the smallest direction (be it normal or
streamwise)).
4.5 FFT calculation
The FFTW library (see (Frigo and Johnson, 2005)) is used to calculate the FFTs. This is a departure
from the original sequential code that implemented a form of the Temperton algorithm for FFTs. Fastest
Fourier Transform in the West (FFTW) is a high-performance library released under the GNU General
Public Licence
. It uses the plan type of interface which has set the standard for other proprietary
implementations (like Intel MKL or CuFFT from Nvidia). Because this interface is standard, use of
other libraries is straightforward. For the current implementation the ability to do batch FFTs was used
(batching n
i
/nprocs 2D FFT planes for each back and forth transition between the spectral and the
physical space). The use of the FFTW library also released the code from several restrictions imposed
by the Temperton algorithm(limits on the values available to create the computational grids, for instance).
Although not used in the present case, FFTW is also able to perform parallel FFTs, either in shared
memory machines (using pthreads or Open Multi-Processing (OpenMP)) or in distributed machines,
using MPI.
4.6 Communications
As stated previously, a choice was made to use the MPI message passing paradigm to deal with
the inter-process communication part of the code. MPI is an open-standard born from the need of
enforcing portability between different parallel architectures. It denes a set of functions (data exchange,
global operations and synchronisation) and data types, leaving out implementation details. This ensures
that the same code can be re-compiled against different MPI library implementations, without code
modications, and so ensuring, where available, the best performance for each architecture.
One of the principal factors impacting the parallelisation of pseudo-spectral DNS codes is the need
to perform all to all communications, in order to transpose the data between the two different partitions.
Two approaches were implemented. Firstly a communication table was created ensuring an exchange
See http://www.fftw.org
4.7 I/O 63
in n
proc
1 execution steps. Later the MPI_ALLTOALL function was used. Considering that no signif-
icant differences in performance were detected, this last form was retained. The general layout of the
exchange subroutine has three steps:
1 Pack data into exchange buffer
2 Make exchange with MPI_ALLTOALL
3 Unpack the data
The data packing routine is responsible for a performance penalty due to the index re-arrangement
(from i, j, k to j, k, i and vice-versa). This re-arrangement promotes large strides when accessing the
arrays, trashing the cache. To minimise this, the data is assembled in a cache friendly way in the send
buffer, so the array unpacking operation has a continuous memory access pattern.
Other operations used in the code are MPI_Broadcast (for sending initial parameters from the master
to all processors), MPI_Allreduce (to do global sums and obtain global maximum) and MPI_Barrier (for
synchronisation purposes).
4.7 I/O
DNS codes are I/O intensive. The amount of needed storage space per timestep grows three times
with the mesh size (for each of the three velocity components). For instance, a 100 million point mesh,
in single precision, needs around 1.2 Gigabyte per timestep for a restart. A 500 million mesh around
6 Gb/timestep. Compounding this problem is the need to make higher resolution simulations or extend
the physical domain further, necessarily entailing an increase in grid points. Another motivation for
le space increase is the need to save a great number of these elds for posterior post-processing,
either statistic, visual or using them in conjunction with other methods (particle tracking, vortex tracking,
structure educing, etc. . . ). This delineates two types of requirements: fast I/O during the run and a
manageable approach to properly archive the resulting les.
I/O is, by nature, one of the slowest operations that can be found in computing systems. This problem
escalates when working in parallel machines. A small resume of the various possible strategies follows
and their relevance to the case at hand.
4.7.1 Only master writes
Figure 4.5: Master writes: all send to master and master writes. This approach is simple, portable but low
performance and must account for possible memory limitations on master node.
In this approach all processes communicate their data to the master node which writes the le to
disk (see g. (4.5)). It has the advantage of being fairly simple to implement and relatively fool proof on
deployment, since there is no special requirement on the underlying lesystem. It is not, however, very
efcient and so, it is very time consuming. The process itself is serialised (each process communicates
its data in turns), and care must be taken for the master to write the le in consecutive chunks. If this
last strategy is not implemented the master must have enough memory space to hold the full array to
be written. If the simulation is not running in a Symmetric Multiprocessing (SMP) machine (where all
processes see the same lesystem), or no parallel lesystem is available, this approach can be used as
a last resort I/O strategy.
4.7.2 All write
Figure 4.6: All write: all nodes can write a partial le on their local (or shared) I/O space. This approach can result
in an unmanageable number of les and is of difcult portability.
An alternative to the previous strategy is having all processes writing their data to disk (see g. (4.6)).
This, although innerly parallel, can have three serious problems:
1. If the lesystem is shared between the nodes saturation problems can arise when all processes
try to do I/O;
2. The number of generated les can be too large to be manageable;
3. Difculties can arise when trying to restart from a different number of processes
Because of these reasons this approach wasnt implemented in the current code.
4.7.3 Use Message Passing Interface Input/Output (MPI-IO)
Figure 4.7: MPI-IO: with the presence of a parallel lesystem or special tuned NFS shared I/O space, MPI
functions can be used to do I/O.
The MPI-2 standard species several functions to execute parallel I/O. It also species ways to ac-
cess those les, so collisions between processes are avoided. Using MPI-IO is then the recommend
4.8 Parallel performance 65
approach for doing I/O in a parallel context. Several implementations take advantage of specic ca-
pabilities of the underlying lesystem to optimise the I/O operations. Because there is a provision to
dynamically adjust the scope of action of each process when writing to the le, restarts can be trans-
parently done from different numbers of processes. MPI-IO does, however, place some restriction on
the user: if not running on the same machine (an SMP for instance) he must have available a parallel
or networked lesystem. If the lesystem is of the NFS type, it has also to be congured with certain
options, something which can degrade its performance.
Albeit these restrictions, MPI-IO is the preferred parallel I/O method and is used in the current code
whenever possible.
4.7.4 Variations and improvements
Figure 4.8: A worthwhile variation of MPI-IO, for very large simulations in big systems is the use of MPI-IO with
dedicated I/O nodes.
A possible variation on the previous strategy is the use of dedicated I/O nodes (see g. (4.8)). This
ofoads the I/O burden from the computing nodes, which can keep calculating during I/O operations.
For simulations using a great number of processes (> 1000) this becomes a requirement. A successfully
example of this use is given by Hoyas and Jimenez (2006) for their channel ow simulation.
4.8 Parallel performance
The two main indicators of a parallel program performance are its speedup, i.e., the time ratio be-
tween a sequential and a parallel execution (eq. (4.3)), and its efciency, eq. (4.4).
S =
T
S
T
N
(4.3)
E
tot
n
=
T
S
NT
N
(4.4)
T
S
is, in expressions eq. (4.3) and eq. (4.4), the wallclock time of a sequential execution and T
N
that
corresponding to the execution using N processors. It must be noted that T
S
does not necessarily
equate to a sequential run of the parallel implementation but should generally use the best sequential
implementation available.
Parallel speedup measures the acceleration (in time) obtained from running the code in parallel (in
N processors) vs. its sequential counterpart. The parallel efciency is a relative measure to gauge how
well the parallel run nears its theoretical potential.
Two motivations underlay the development of parallel codes: obtain results earlier or to solve larger
problems, out of reach with sequential resources. Accordingly, two different ways exist to look at parallel
speedup: Amdahls and Gustafsons law. Their description follows with a general comment on the factors
affecting parallel speedup and efciency.
4.8.1 Speedup - Amdahls law
Amdahls law measures the speedup for constant size problems when the number of available pro-
cesses increases. Formulated by Gene Amdahl while working at IBM (Amdahl (1967)), it assumes that
programs can be divided into sequential and parallel parts. While the time spent in the parallel part
diminishes with increasing number of processors, the time spent in the sequential part stays the same,
imposing a limit to the speedup growth with increasing number of processors. Amdahls law is expressed
by eq. (4.5)
S =
1
(1 P) + P/N
(4.5)
where speedup (S), parallel fraction (P) and number of processors (N). It is plain that Amdahls Law
% of parallel part:
100.0%
99.9%
99.0%
90.0%
80.0%
50.0%
S
p
e
e
d
u
p
1
2
4
8
16
32
64
128
256
512
1024
2048
N
1 2 4 8
1
6
3
2
6
4
1
2
8
2
5
6
5
1
2
1
0
2
4
2
0
4
8
Figure 4.9: Amdahls law
promises a dark cloud for almost all parallel codes. Unless they are of the embarrassingly parallel
type, where no sequential part exists, a limit to their speedup can be rapidly reached, as can be seen in
g. (4.9). It is easy to see that even with a 99.9% parallel code the model predicts a deviation from the
perfect speedup from 64 processors onward.
Communication costs can also be accounted into the sequential cost of the program. Such is of
particular relevance for the pseudo-spectral DNS code dealt here: the high costs of communication
associated with the data transposition quickly shifts the speedup growth from the linear zone. When
running in computing facilities with a computational cost economy in place (generally using CPU/hour
budgets), it may be more protable to use less processors to achieve a higher number of timesteps
under the same project budget. Fortunately Amdahls Law is only half of the story.
4.8.2 Scalability - Gustafson law
Although Amdahl predicts a limit for the speedup growth in xed size problems, another narrative is
obtained when the problem size grows. This is a scalability concern best expressed by Gustafsons Law,
which expresses the scaled speedup (SS) (Gustafson (1988)):
SS = (1 P) + NP (4.6)
Gustafson compares the gain between using N machines to run a parallel program with its sequential
counterpart. It exposes scalability behaviour through the variation of the sequential cost (assumed
constant in Amdahl), with increasing problem size and processors available (refer to g. (4.10)). In other
words, if the total running time is decomposed into a parallel and a sequential fraction, it is assumed that
their relative cost is kept constant while increasing the problem size. As can be seen by this last gure,
the sequential fraction cost is mitigated with increasing N, and the problem size that can be solved grows
linearly with the number of processors. This means that even with a parallel part of only 50%, as seen
in the gure, it can still be interesting to use the parallel code to solve bigger problems, not achievable
any other way.
% of parallel part:
100.0%
80%
50%
S
c
a
l
e
d

S
p
e
e
d
u
p
1
128
256
512
2048
N
1
1
2
8
2
5
6
5
1
2
2
0
4
8
Figure 4.10: Gustafsons law
4.8.3 Additional considerations on parallel performance
Parallel performance depends on several factors. A small review follows:
Numerical efciency: sequential algorithms do not necessarily transpose efciently to a parallel
environment. One paradigmatic example is that of the Thomas algorithm for solving tridiagonal
matrices. The Thomas algorithm is a simplied form of the Gaussian elimination and can be
summarised in a two step algorithm: a rst sweep from top to bottom eliminates the lower diagonal
and then a backward sweep from bottom to top obtains the solution. Transporting this algorithm
to a parallel machine results in the pipelined Thomas algorithm which, while it progresses through
the matrix keeps several processors idle throughout the whole, becoming inefcient (see Povitsky
(1998)).
Parallel efciency: this sums the impact of the communication phase of the program. Generally
speaking, while the program is doing communication it is not calculating, but wallclock time still
increases. Calculation and communication overlap can sometimes be used but such is, in prac-
tice, difcult to achieve: algorithm contexts where it can be implemented are rare and the right
combination of MPI library and hardware is still necessary for its full realisation. Parallel efciency
can be further dissected into:
Latency, i.e., the overheads paid because of transmission preparation;
Data transfer, i.e., the time spent in the transaction;
Floating point operations, i.e., the ratio of time spent doing oating point operations.
Load balancing: a balanced distribution of the computational load is of utmost importance for
parallel performance. Unbalanced programs are slower because they promote some processors
to become idle while waiting for their counterparts to nish their work.
4.9 Final notes on porting sequential codes
A few nal notes are in order about porting sequential codes to parallel architectures. These are in
fact common to all faced with software legacy issues. The original sequential code was programmed
using the F77 standard of the Fortran language, making extensive use of common blocks. Although an
initial parallel version was implemented using this version of Fortran, several issues related with lack
of exibility, deployment and code extension were soon faced. Fortran has been an evolving standard,
adapting to the needs of its users, most belonging to the numerical scientic community and especially
the working in high performance computing contexts. This problem is transversal because of the many
legacy codes still running worldwide.
Fortran 2003 is the last version of the Fortran standard (version 2008 is being discussed right now).
It adds better support for Object Oriented (OO) programming but, more important, has better support for
C language interaction (allowing the memory optimisation explained previously). Changing from Fortran
77 was deemed necessary, especially because:
Ability to change the grid size by user input, without need to recompile the code (from Fortran 90
onward this is guaranteed because it becomes possible to dynamically allocate/deallocate arrays);
Elimination of COMMON blocks;
Possibility of code extension easier due to OO support;
Better interaction with C language expands the scope for exploring; other libraries and hardware
(the recent developments in GPU, for instance).
Code management is many times overlooked in scientic codes. This is a management failure
inasmuch as these codes can have very long lives, being extended and updated by different generations
4.9 Final notes on porting sequential codes 69
of engineers and scientists. A solid and clean base of work is important in this respect, going together
with code good documentation and practices.
CHAPTER 5
Post-processing tools
This chapter describes some of the post-processing tools developed to achieve the goals at hand,
namely the vortex tracking algorithm and the turbulent/non-turbulent interface detection scheme. A
small foreword is written on the account of post processing tools for turbulent ows.
5.1 Post-processing tools, a foreword
An hidden problem of large simulations is the size of the data being generated. For instance, the
simulation of a channel ow by Hoyas and Jimenez (2006) originated up to 25TB of data. Retrieving
meaningful results from this data can be as daunting as generating it. As the data can be re-used for a
large number of different studies some research groups start to opt for post-processing clusters. In these,
the data storage is combined with post-processing software and hardware capabilities. This allows
inquiries and data manipulation Li et al. (2008). Another strategy consists using distributed computing,
collecting unused resources available in laboratory workstations through the use of High Throughput
Computing (HTC) managing software like Condor.
For data storage and re-use public servers interest in public servers and portals is also increasing.
They are used to make the data available to all the research community. Such type of initiatives have
been pursued by individual research groups (like Hoyas and Jimenez (2006)) or through public portals,
as the iCFD database. These later also try to answer questions pertaining results crosscheck, a concern
that has been raising among the scientic community at large during the last years.
72 Post-processing tools
5.2 Vortex identication
This subsection details the algorithm implemented to identify and characterise the intense vorticity
structures (IVS) present in turbulent ows. Several identication and tracking algorithms have been
used throughout the literature, emphasising different aspects of the phenomena of interest as described
in chapter 2. Also as described in the same introductory chapter, the IVS are coherent structures, of
tubular shape, with high vorticity content, also known as vortex tubes or, more colloquially, as worms.
Although these structures ll up a small amount of the ow total volume, they account for most of the
vorticity present. Another interesting aspect to note is that the highest vorticity areas seem to self-
organise in tubes or ribbons (see Jimnez et al. (1993) and g. (5.1)).
Figure 5.1: The left image shows vorticity streamlines conditioned for || > 60 in an HIT simulation. The
organisation of vorticity streamlines into bundles is readily apparent. On the right, vorticity iso-contour surfaces for
the same threshold can be observed, making the tube and ribbon like structures very noticeable.
As in Jimnez et al. (1993); Jimnez and Wray (1998), worms are dened as having an intense
vorticity level, i.e., those belonging to the 1% of the ow volume with highest vorticity norm (i.e., above
ivs
). Similarly, weak vorticity is dened as [[ <
. Those in between these levels are considered as

belonging to the background vorticity (see g. (5.2)).
Figure 5.2: Vorticity levels
is dened by eq. (5.1):
2
= / (5.1)
The vortex tracking algorithmimplemented is similar to Jimnez et al. (1993) with some modications.
An homogeneous isotropic turbulence simulation was used to verify the algorithm implementation. In
5.2 Vortex identication 73
this simulation the particular value of
ivs
= 2.7
and can be obtained from the histograms in Fig. 7.1

in chapter 7. For comparison Jimnez et al. (1993) obtained
ivs
= 3.1
at similar Reynolds numbers.

As a pre-processing step in the tracking algorithm, all the identied points i.e., points with vorticity
above the dened threshold >
ivs
are listed and ordered. Each axis is then identied according to
the following steps:
i) choose the rst non assigned point;
ii) choose the nearest grid plane intersected by the direction of . The four points in this plane nearest
to the intersection are chosen as candidates (blue in g. (5.3);
Figure 5.3: Vortex ID stencil: Red circle marks the current selected point. Blue circles mark the next 4 candidates.
Straight lines signal the intersection plane of the vorticity vector from the current selected point.
iii) choose the non-assigned point with the highest value of
2
;
iv) stop when all listed points have been assigned.
The algorithm stops when all points have been assigned. In a second post-processing phase all IVS
points belonging to an axis with less than 20 points are discarded. A modication introduced here
absent from the tracking described in Jimnez et al. (1993) is the elimination of all the clustered axis
points (i.e., points from two parallel axis), leaving only points from the axis with the strongest enstrophy.
The worm radius and circulation as function of axis position are calculated from the vorticity prole
along the axis. This prole is obtained using the following steps:
i) a plane normal to the direction of the vorticity vector is dened for each identied point;
ii) the vorticity can be interpolated into any given point on this plane, using the least-squares method.
In the present work n = 6 radial bins centred on the axis, with a radius equal to nx, and with m = 8
points per bin, with an angular interval equal to 2/n were used;
iii) to compensate for the noise induced by the discretisation, the value obtained for the radial distribu-
tion is ltered over triples of consecutive axial locations using a [1/4, 1/2, 1/4] mask.
The worm radius R is calculated by an iterative t of the axial distribution to =
0
e
r
2
/R
2
, where
0
is the value of vorticity at the axis point. All points where the iterative process was unable to converge
were discarded, like those with a very big radius R > 30, as in Jimnez and Wray (1998). Fig. (5.4)
shows an IVS immersed in the turbulent ow of an HIT simulation showcasing the distribution of the
radial bins to calculate the prole.
Figure 5.4: IVS educing sketch. Figure shows || = 100 iso-contour surfaces with one IVS educed (vorticity
vectors in yellow/green). The orange vectors also represent vorticity and are distributed along the radial bins
mentioned in the algorithm description.
The tangential (or azimuthal) velocity was computed using two different methods. The rst method
assumes a Gaussian distribution for the vorticity in each IVS (r) =
0
e
r
2
/R
2
. Using the relation
_
u
dl =
_ _
ndS with the assumed vorticity distribution one gets the expression for the tangential
velocity u
0
= 0.316
0
R.
In the second method, the tangential velocity is determined directly from the velocity ow eld. A
plane normal to the
0
axis is determined and the ow eld vectors are projected into it. The tangential
velocity is then obtained at each radial position using an azimuthal averaging procedure similar to the
one described above. The two methods give similar results for u
0
. In the present work the second
method was used because it requires no assumption on the shape of the vorticity distribution.
The algorithm implementation verication is presented in chapter 7.
5.3 Interface detection and statistics
5.3.1 Detection
The detection of the turbulent/non-turbulent interface is not a straightforward subject. Theoretically,
this boundary should be detected by just checking for the absence/presence of . Unfortunately, the
existence of background noise (introduced by the experimental measuring devices in the laboratory or
the numerical noise present in the DNS/LES simulation) precludes an exact knowledge of its location.
All T/NT interface detection methods are then approximate and generally based in a threshold criteria of
5.3 Interface detection and statistics 75
some property of the ow, e.g., vorticity, concentration of passive scalar or a velocity component - for a
brief discussion on available methods see Anand et al. (2009).
For the work presented on this thesis, the interface detection algorithm is the same that was rst
described by Bisset et al. (2002) and later also used by Westerweel et al. (2005, 2009); da Silva and
Pereira (2008) and da Silva (2009). The criteria used is based on the vorticity norm [[ =
i
. Like in
Bisset et al. (2002) the value of [[ = 0.7U
1
/H was found to best delineate the vortical regions:
_
_
_
[[ > 0.7U
1
/H turbulent
otherwise non turbulent
(5.2)
The interface is obtained detecting the threshold in the vorticity eld starting from afar and progressing
in direction of the jet core. Notice that the interface is processed in 2D X0Y planes and the line thus
obtained will necessarily encompass some irrotational ow (refer to g. (5.5)).
Figure 5.5: Interface diagram. Turbulent ow is depicted with gray, the interface line is dashed. Global and local
reference axis are also portrayed.
5.3.2 Conditional statistics
Conditional statistics allowa better understanding of the behaviour of the owproperties in the vicinity
of particular events. Several of these statistics were done, conditioned to the interface distance, with
the purpose of furthering the knowledge about the physical mechanisms of the T/NT interface. The
approach taken is similar to that present in Bisset et al. (2002); Westerweel et al. (2005, 2009); da Silva
and Pereira (2008); da Silva (2009).
For these, a local reference axis Y
I
was dened, with origin in the interface (see g. (5.5)). Areas of
enclosed irrotational uid, trapped amid the turbulent zone, or turbulent globs isolated in the irrotational
zone, were excluded from the statistics, in the same manner as in e.g., Westerweel et al. (2009).
If statistics of specic points were desired (i.e., like those of the identied worm axis, which will be
mentioned later), the distance Y
I
was measured along a vertical line until intersection with the interface
was obtained (see g. (5.6)). The correct placement of the referential is achieved using a linear inter-
polation. The local axis positive direction is always toward the centre of the jet and the T/NT interface
is at y
I
= 0, while the irrotational and turbulent regions are dened by y
I
< 0 and y
I
> 0, respectively.
These conditional statistics were denoted by
I
. Tests were done comparing the conditional statistics
Figure 5.6: Conditional statistics of isolated interest points.
obtained using a local axis lined up with the interface tangent against the vertical option show in g.
(5.5). No relevant differences were noticed.
<||>
I
<|
x
|>
I
<|
y
|>
I
<|
z
|>
I
0
1
2
3
4
y
I
/
0 25 50 75
Figure 5.7: Mean conditional proles of ||
I
, |
x
|
I
, |
y
|
I
and |
z
|
I
(all normalised by U
1
/H)
Fig. (5.7) shows average vorticity proles conditioned to the interface distance. It is clear from these
that a sharp jump happens in the vicinity of the T/NT interface. These results are further explored in
chapters 6 and 7.
Part III
Results and Discussion
CHAPTER 6
The role of coherent vortices near the turbulent/non-turbulent (T/NT)
interface in a plane jet
This chapter focus on the role the coherent vortices near the T/NT interface, either IVS or LVS. It
presents, among others, results on the T/NT interface thickness and explains the presence of viscous
dissipation in the irrotational region adjacent to the interface. The results here presented were
published in da Silva and dos Reis (2011).
6.1 Introduction
Turbulent entrainment is one of the most important features observed in free shear ows such as mix-
ing layers, wakes and jets and governs some of the most important characteristics of these ows such
as their spreading and mixing rates. Turbulent entrainment takes place across a sharply contorted inter-
face - the turbulent/non-turbulent (T/NT) interface - that divides the turbulent (T) from the non-turbulent
(NT) or irrotational ow regions (Corrsin and Kistler (1955)). Understanding of the physical mechanisms
taking place at the T/NT interface is important to many natural and engineering ows since important
exchanges of mass, momentum and passive or active scalar quantities e.g. heat, take place across the
T/NT interface.
The important role played by the coherent vortices in the context of the turbulent entrainment mech-
anism has been recognised long time ago (Townsend (1966)). Past studies assumed that turbulent
entrainment is mainly caused by large-scale engulng motions induced by the coherent vortices in the
vicinity of the T/NT interface (Townsend (1966)). It is clear that these structures drive many of the impor-
tant process that take place near a T/NT interface (Bisset et al. (2002); Hunt et al. (2006, 2009, 2008)).
Specically it is thought that the entrainment rate is somehow imposed by the motion of these large
80 The role of coherent vortices near the turbulent/non-turbulent (T/NT) interface in a plane jet
scale eddies near this interface.
However it has been shown recently that the entrainment is mainly caused by small scale eddy
motions (nibbling) acting on the T/NT interface (Mathew and Basu (2002); Westerweel et al. (2005)),
as suggested originally by Corrsin and Kistler (1955). This recent discovery motivated a number of
investigations on the mechanisms behind the large and small scale features responsible for the turbulent
entrainment.
Bisset et al. (2002) observed that the vorticity components undergo a sharp jump at the T/NT inter-
face. Westerweel et al. (2005, 2009) observed the existence of a jump in the streamwise velocity and
passive scalar eld across this interface and conrmed that the engulfment is not the dominating pro-
cess for the entrainment. They also analysed the dynamics of the ow relative to the T/NT interface and
investigated the eddy viscosity near that location. Holzner et al. (2007, 2008) analysed the dynamics
of the enstrophy and strain near a T/NT interface generated by an oscillating grid. They observed the
existence of an intense kinetic energy dissipation outside the turbulent region near the T/NT interface
and also found out that the net effect of viscosity at that location is to cause an increase of the total
enstrophy. Subsequently it was observed by Holzner et al. (2008) and da Silva and Pereira (2007a) that
it is the viscous diffusion that is responsible for this positive viscous contribution to the enstrophy while
the enstrophy viscous dissipation remains negative across the jet, as expected. Recently Holzner et al.
(2009) decomposed the entrainment velocity into two terms: a viscous and an inviscid term. It was ob-
served that the viscous term dominates the entrainment velocity and that its magnitude is comparable to
the magnitude of the local Kolmogorov velocity. The universal small scale features such as the geometry
of the dissipation, the geometry of the straining (or deformation) of the uid elements, and the behaviour
of the invariants of the velocity gradient tensor - Q and R - during the entrainment process was studied
by da Silva and Pereira (2008, 2009). They observed that the characteristic tear drop shape in the
(R, Q) map is not formed at the T/NT interface, but that it needs about one Taylor micro-scale into the
turbulent region in order to form completely. Finally, the challenges faced by the subgrid-scale models
near the T/NT interface were analysed by da Silva (2009). The classical subgrid-scale models may need
particular modications near the edge of a jet if ne capturing of the Reynolds stresses at this location
is needed e.g. in order to predict accurately the mixing rates near the T/NT interface of a jet.
However many aspects associated with the characteristics and dynamics of the T/NT interface re-
main unknown. One of these aspects consists on the existence of a viscous super-layer where viscous
processes dominate, as suggested long time ago by Corrsin and Kistler (1955). A related problem con-
cerns the thickness of the sharp vorticity jump observed at the T/NT interface. Reynolds (1972) predicted
that a vorticity jump is to be expected to occur at the T/NT interface. However the values observed for
the thickness of this jump vary in the literature e.g. the thickness of the turbulent front generated from
an oscillating grid is of the order of the Kolmogorov micro-scale (Holzner et al. (2007)) whereas the
experimental results from round jets by Westerweel et al. (2005, 2009) and the numerical simulations of
plane jets by da Silva and Pereira (2008, 2009) show clearly that the thickness of the vorticity jump at
the T/NT interface is of the order of the Taylor micro-scale.
Another issue concerns the role of the coherent vortices near the T/NT interface in jets and wakes,
6.2 Direct numerical simulation of a turbulent plane jet 81
as opposed to mixing layers (Westerweel et al. (2009)). Recent works addressed the ow motion in-
duced by the presence of these structures near the jet edge. Bisset et al. (2002) studied the shape
of the streamlines caused by the large-scale motions near the T/NT interface. They observed that the
streamlines of the entrainment wind only cross the T/NT interface where they are normal to the interface
surface. The characteristics and dynamics of the shear layers bounding the turbulent and the irrotational
ow regions in shear ows are determined by the presence of the nearby vortices. In this context many
useful estimates were derived using linearised calculations by Hunt et al. (2006, 2009, 2008). Recently
Hunt et al. (2010) discussed the structure of the shear layers between adjacent eddies. The process
of nibbling is for the rst time described as a two scale process whereby the vortices that lie within a
shear layer with thickness equal to the Taylor micro-scale undergo local straining causing their thickness
to be of the order of the Kolmogorov micro-scale. However, many aspects of the dynamics of these
structures are still unclear. If the entrainment rate is imposed by the large scales but ultimately is felt at
the small scale nibbling motions how exactly does this process occur i.e. how do the large and small
scale motions interact to cause the turbulent entrainment?
The goal of the present work is to shed light into the role of the coherent vortices in a turbulent jet near
the T/NT interface. For this purpose a direct numerical simulation (DNS) of a turbulent plane jet will be
used. The coherent vortices of jets share many common features with the structures commonly found in
other free shear layers such as azimuthal tube vortices (generated by the Kelvin-Helmholtz instability in
the case of jets and mixing layers), and the existence of streamwise secondary vortices between these
primary structures whose "foot prints" are still discernible at the far eld fully developed turbulent state.
Therefore, it is expected that the results obtained in the present study display some universal qualitative
features of the turbulent entrainment that exist in free shear ows in general.
This chapter is organised as follows. Section II contains the description of the plane jet DNS used
in the present work and the procedure used here to obtain the conditional statistics in relation to the
T/NT interface and to detect the IVS. Section III describes the results. These concern (i) the thickness
of the vorticity jump near the T/NT interface, (ii) the orientation of the coherent vortices in respect to the
surface dening the T/NT interface, (iii) the kinetic energy viscous dissipation near the jet edge and, (iv)
the anomalous enstrophy viscous diffusion near the T/NT interface. The work ends with an overview of
the main results and conclusions.
6.2 Direct numerical simulation of a turbulent plane jet
In the present work a direct numerical simulation (DNS) of a turbulent plane jet is used to analyse
the effect of the coherent structures near the turbulent/non-turbulent interface. This DNS is described in
detail in da Silva and Pereira (2008); da Silva (2009) and da Silva and Taveira (2010).
It is noteworthy that the present simulation is mathematically identical to the plane wake used in
Bisset et al. (2002), since temporally evolving wakes and planar jets differ mainly in the shape of the
initial mean velocity prole: the maximum mean velocity in a jet is located at the centreline while in
a wake the minimum velocity is located at the centreline. Consequently in the self-similar regime the
shape and magnitude of the Reynolds stresses proles in the two ows are similar as can be attested
by comparing those present in da Silva and Pereira (2008) and Bisset et al. (2002).
The Navier-Stokes solver uses a pseudo-spectral scheme for spatial discretisation, and a 3rd order,
3 steps, Runge-Kutta scheme for temporal advancement. The grid is isotropic (x = y = z),
and the number of collocation points along the streamwise (x) normal (y) and spanwise (z) directions
is equal to (N
x
N
y
N
z
) = (256 384 256). The extent of the computational domain attains
(L
x
, L
y
, L
z
) = (4H, 6H, 4H), where H is the initial jet slot width.
The initial condition consists of a mean velocity prole to which a three-component velocity uctuating
"spectral noise" was superimposed
. Specically, the mean velocity prole is given by,

u(x, y, z) =
U
1
2
_
1 + tanh
_
H
4
0
_
1
2[y[
H
__ _
, (6.1)
where U
1
is the mean velocity at the jet centreline and
0
is the initial momentum thickness. The non-
dimensional ratio between the inlet slot-width and the initial momentum thickness was set to H/
0
= 35.
Similar mean velocity proles were used in many other simulations of planar jets e.g., Stanley et al.
(2002).
The simulations are halted before the effect of the boundary conditions can be observed in the jet
statistics e.g., the Reynolds stresses. Extensive validation tests were undertaken for this simulation
and the results showed that the present DNS is accurate at the large and small scales of motion and
representative of a fully developed turbulent plane jet (see reference da Silva and Pereira (2008) for
details). The self-similar regime occurs when the second order moments at several different times
collapses, and is obtained at T/T
ref
20, where T
ref
= H/(2U
x
) and U
x
is the initial velocity at the
centreline by which time the jet half width is equal to
0.5
/H = 0.78. At the self-similar regime the
Reynolds number based on the Taylor micro-scale
2
=
u
2
_
/
_
(u/x)
2
_
, and on the root-mean-
square of the streamwise velocity u
u
2
_
1/2
is equal to Re
= u
/ 120 across the jet shear layer.

The resolution is x/ 3 across the jet shear layer. This is slightly less than in isotropic turbulence
but is well in line with the resolution used in DNS of free shear ows. The kinetic energy and enstrophy
spectra da Silva and Pereira (2008) conrm that the resolution is adequate.
The ow coherent structures from the plane jet DNS are qualitatively similar to many previous di-
rect numerical simulations of turbulent plane jets e.g. Stanley et al. (2002). Figures 6.1 (a,b) show
iso-surfaces of Q > 0 and pressure, respectively, at the far eld (self-similar) regime, where Q =
i
/4+S
ij
S
ij
/2, where S
ij
= 1/2 (u
i
/x
j
+ u
j
/x
i
) is the rate-of-strain tensor, and
i
is the vorticity
vector. The lowpressure iso-surfaces highlight the big rollers which are remnants of the Kelvin-Helmholtz
vortices generated during the transition to turbulence which represent in the typical LVS from a turbulent
plane jet. The iso-surfaces of Q > 0 highlight smaller, more intense structures, with less signs of a par-
ticular spatial orientation - except for the streamwise vortices near the jet edges. As expected the IVS
are closer to the structures dened by Q > 0 than to the structures dened by regions of low pressure
(see below), but there is of course some spatial overlap between these structures.
See sec. (3.6).

Figure 6.1: Iso-surfaces of Q = 30(U
1
/H)
2
(a) and pressure p = 0.1(U
2
1
) (b) at the self-similar regime
(T/T
ref
22). Note that the gures do not show the total extent of the lateral domain.
6.2.1 Detection of the T/NT interface: conditional statistics in relation to the distance from the
T/NT interface
The T/NT interface is detected using a procedure based on the vorticity norm threshold described
earlier (chapter 5) and g. (5.7) is repeated here from the same chapter as g. (6.2). It shows conditional
mean proles of [
x
[
I
, [
y
[
I
and [
z
[
I
in relation to the distance from the T/NT interface, showing
a sharp jump across the T/NT interface with thickness roughly equal to
20. For the present

simulation inside the turbulent region there are 20, where is the Taylor micro-scale, therefore this
jump is equal to the
, in agreement with other experimental and numerical works (Westerweel

et al. (2005, 2009); Mellado et al. (2009)).
<||>
I
<|
x
|>
I
<|
y
|>
I
<|
z
|>
I
0
1
2
3
4
y
I
/
0 25 50 75
Figure 6.2: Mean conditional proles of ||
I
, |
x
|
I
, |
y
|
I
and |
z
|
I
1
/H)
6.2.2 The coherent vortices near the T/NT interface
In turbulent ows the regions of concentrated vorticity consist of either (i) vortical tube like structures
(vortex tubes) or in (ii) structures exhibiting a sheet like shape (vortex sheets) Horiuti (2001). Whereas
in the tube like structures vorticity and strain are comparable, the sheet like structures are dominated
by vorticity with relatively low levels of strain. The sheet like structures exhibit smaller lifetimes than
the tube structures and therefore it is likely that the tubes will have more possibilities of inuencing the
dynamics of the ow near the T/NT interface than the relatively short lived sheet structures Horiuti and
Takagi (2005).
For this reason this paper is focused on the tube structures and these structures were divided into
two classes: intense vortical structures (IVS) and large scale vortical structures (LVS).
The IVS are dened as structures with particularly strong vorticity i.e. structures comprised of ow
points with vorticity greater than a particularly high vorticity threshold. Jimnez et al. (1993); Jimnez
and Wray (1998) dened this vorticity threshold as equal to the vorticity dening the 1% of ow points
with the highest vorticity. In isotropic turbulence the IVS are the well known worms described by Siggia
(1981). The IVS are similar in many different ows such as mixing layers, wakes, jets and also boundary
layers, ducts and channel ows (Tanahashi et al. (2001); Mouri et al. (2007); Kang et al. (2008); Gana-
pathisubramani et al. (2008)). For instance the IVS in these ows do not exhibit any particular spatial
orientation and their vortex core radius is R
ivs
/ 4 5 where is the Kolmogorov micro-scale.
It was dened by LVS all the remaining vortex structures i.e. structures of concentrated vorticity with
tubular shape with vorticity smaller than the IVS. The LVS are the largest vortical structures which are
present in a particular ow. Often originating in the particular instabilities from that ow, their characteris-
tics such as the vortex core radius, azimuthal velocity and life time are deeply related to these processes
and therefore are quite different from ow to ow. However the dynamics of these structures share some
common features with LVS from other turbulent ows e.g. they consist in structures with roughly tubular
shape and are approximately governed by the same simple inviscid laws.
Although there is some arbitrariness in these denitions the separation between the two classes of
vortices is important since the two types of structures are fundamentally different.
Figure 6.3 displays both the large-scale vortical structures (LVS) and the intense vorticity structures
(IVS) near the T/NT interface for the present turbulent plane jet simulation. The ow coherent structures
from the simulated jet are similar to many previous direct numerical simulations of turbulent plane jets
e.g. da Silva and Mtais (2002a); Stanley et al. (2002). The LVS are identied here using low pressure
iso-surfaces. The LVS (white) consist in the classical big rollers and in the streamwise vortices that
are known to exist in a turbulent plane jet. The rollers in particular can be seen as remnants of the
Kelvin-Helmholtz vortices generated during the transition to turbulence in the jet while the streamwise
vortices correspond to the secondary instabilities. A very large Kelvin-Helmholtz roller can be seen
at the right side of the gure, extending through the whole height of the gure with several legs from
streamwise vortices that are seen to be emerging from this roller. The observed radius of the bigger LVS
i.e. the rollers is close to the Taylor micro-scale R
lsvs
. Notice that the LVS not only exhibit some
level of preferential spatial orientation close to the T/NT interface but also are clearly responsible for the
characteristic size of the convolutions observed on the T/NT interface i.e. the observed convolutions of
the T/NT interface are clearly associated with the presence of a nearby LVS underneath. Consequently,
the average length of these convolutions is deeply inuenced by the geometry of the LVS. Specically,
the radius of these structures is clearly proportional to the average length of the convolutions of the T/NT
interface, in agreement with previous observations (e.g. Townsend (1966)).
The IVS (yellow) are identied here using a similar tracking procedure as the one developed in
Jimnez et al. (1993) and Jimnez and Wray (1998) where the IVS are dened as consisting of points
where the vorticity = [ [ is above a particular vorticity threshold
tr
dening the points with highest
enstrophy that are contained in 1% of the total volume. Once the axis points for the IVS were identied
the IVS radius R for any point on the axis is calculated by tting the axial distribution to (r) =
0
e
r
2
/R
2
,
where
0
is the value of vorticity at the axis point. As can be seen in Fig. 6.3 the IVS exhibit a much
smaller radius than the LVS also and a more "random" spatial orientation. The calculated core radius of
the observed IVS is R
ivs
4.6, where is the Kolmogorov micro scale, which agrees with the values
observed in isotropic turbulence (Jimnez et al. (1993); Jimnez and Wray (1998)). Notice however that
there is of course some spatial overlap between the LVS and the IVS since the IVS are extreme events
of LVS. Its interesting to observe that this gure resembles the sketch of the LVS and IVS described in
Westerweel et al. (2009) where the IVS are described as structures surrounding the LVS near the jet
edge.
Following a suggestion from Javier Jimnez (Private Communication) it is argued here that the T/NT
interface in fact consists of (or is made up from) the borders of the coherent vortices sitting at that
location. Notice that vortex sheet structures described in the introduction are also part of the T/NT inter-
face but their effect on the T/NT interface statistics is small due to their smaller life time and coherence
compared to the vortices. For the same reason the geometrical characteristics of the T/NT interface is
expected to be more dominated by the LVS than the IVS simply because, the bigger the radius the bigger
the life time of a given vortex, and thus the bigger his chances of dening the geometry and statistics
of the T/NT interface. In the same vein it is expected that the larger of the LVS e.g. the rollers to have
more inuence than the smaller LVS i.e. the streamwise vortices. However instantaneously all the range
of LVS and IVS of course contribute to the denition and characteristics of the T/NT interface. Figure
6.3 supports this view in that it shows that the geometry of the T/NT interface is more inuenced by the
larger LVS than by the smaller LVS or the IVS lying underneath its surface.
In order to show this Figs. 6.4 (a) and (b) are used to show that the IVS have only a limited inuence
on the characteristics of the T/NT interface i.e. something else has to exist near the jet edge in order
to cause the observed T/NT interface characteristics such as its thickness, that as seen, is equal to the
Taylor micro-scale
. Figure 6.4 (a) shows the Gaussian vorticity prole used to t the vorticity
distribution for three different IVS with R/ = 4, 5 and 6 using the method described in Jimnez and
Wray (1998). l
is dened as the distance from the axis of the IVS to the point where the vorticity equals
the threshold used here to dened the T/NT interface e.g. for a IVS with radius equal to R/ = 6, the
radial distance r to which the vorticity (r) is equal to the threshold used to dened the T/NT interface
is l
/ 12. Figure 6.4 (b) shows the conditional prole for l
, where l
is computed for each one of

Figure 6.3: Large-scale vortical structures (LVS) and intense vorticity structures (IVS) near the T/NT interface from
the upper shear layer of the planar turbulent jet (the ow is from left to right): The T/NT interface (translucent
orange) displays a contorted shape dictated by the large scale vortex structures (LVS) underneath its surface. The
LVS are dened by low-pressure iso-surfaces (white), while the IVS (solid yellow iso-surfaces) are dened by
points where the vorticity = | | is above the vorticity threshold
tr
, which is equal to the vorticity of the points
with highest enstrophy that are contained in 1% of the total volume. The IVS structures are depicted at their real
scale, with their radius calculated from the algorithm in sec. (5.2).
6.3 The role of coherent vortices near the turbulent/non-turbulent (T/NT) interface 87
the detected IVS. Inside the region 0 < y
I
/ < 20 the conditional prole of l
is approximately linear,
however the slope of this line is not equal to 1 e.g. l
/y
I
0.2 ,= 1. Therefore, the conclusion is
that the geometry of the T/NT interface is not imposed by the shape of the IVS. On the contrary, this
results suggests that these vortices have a small impact on the determination of the vorticity jump across
the T/NT interface. In line with this, conditional statistics of the number of detected IVS as function to
the distance from the T/NT interface were determined (not shown) and indicate that no IVS axis exist
between the T/NT interface and a distance of roughly 5 into the interior of the turbulent region. This
is not surprising: since the radius of the IVS is of the order of 5 naturally only IVS with radius much
smaller than 5 could t in the region dened by 0 < y
I
/ < 5.
These results agree with the visualisations shown in Fig. 6.3 which further supports the idea that the
T/NT interface is in fact the physical line dened by the borders of the range of vortices near the edge of
the jet, from LVS down to IVS, however the geometry of the T/NT interface is likely to be more affected
by the bigger LVS than to the smaller LVS or the IVS, since the bigger vortices evolve slowly and have
longer lifetimes and thus are able to inuence the T/NT interface in a more permanent way.
6.3 The role of coherent vortices near the turbulent/non-turbulent (T/NT) inter-
face
This paper focuses on the effect of the LVS and IVS on the turbulence dynamics and characteristics
of the ow near the T/NT interface. The next sections address the thickness of the vorticity jump,
orientation of the coherent vortices, energy viscous dissipation, and enstrophy viscous diffusion near
the T/NT interface.
6.3.1 The thickness of the vorticity jump at the T/NT interface
An interesting and long standing issue regarding the T/NT interface has been the magnitude of its
thickness
. In the original work by Corrsin and Kistler (1955) it is postulated that a viscous superlayer
must exist outside the T/NT interface, where the viscous diffusion of vorticity responsible for the growth
of the shear layer takes place. Assuming that this thickness depends only on the viscosity and on
the enstrophy inside the shear layer
, simple dimensional reasoning yields
(,
) =
_
/
=
_
3
/
_
1/4
= . Holzner et al. (2007, 2008) showed that in the T/NT interface originated froman oscillating
grid the thickness of the vorticity jump is of the order of the Kolmogorov micro-scale, however both the
experimental results in round jets from Westerweel et al. (2005, 2009) and the DNS of planar jets by
da Silva and Pereira (2008, 2009) show that the thickness of this jump is equal to the Taylor micro-scale.
A justication for these two apparently conicting results can be given if one realises that the T/NT
interface is largely made up from the coherent vortices sitting near the T/NT interface. Arguably the
T/NT interface is made up both from vorticity from the whole range of eddies near the T/NT interface
and also from some incoherent vorticity that is shed by these vortices as they travel into the T/NT
interface. However, this incoherent vorticity as is well known is short lived i.e. its life time is small
compared to the life time of the coherent vortices, and moreover it is difcult to assign a proper length
l
/
R/=6
R/=5
R/=4
(
r
/
)
0
5
10
15
20
r/
0 5 10 15 20
(a)
<l
w
>
I
/
6
8
10
12
y
I
/
0 25 50
(b)
Figure 6.4: (a) Sketch explaining the determination of the theoretical distance l
; (b) Conditional prole of l
.
scale to these islands of vorticity lost by the eddies. As explained before from all the sea of eddies
dening the T/NT interface the bigger eddies from the class dened here as LVS should be the ones
responsible for the T/NT interface characteristics since they are more permanent, evolve slower than the
other vortices, and moreover are associated with larger lifetimes.
Assuming that the length scale of this vorticity layer caused by the vortices
depends only on the

viscosity and on the strain rate acting on the LVS it can be writen
(, S), where S is the strain

rate acting on the LVS. It can be argued that the magnitude of the strain rate acting on a LVS near the
jet edge is S u
/L
11
, where u
is the uctuating velocity eld and L

11
is the integral scale of turbulence
(Hunt et al. (2006, 2010)). If the Reynolds number associated with the integral scale is Re
0
= u
L
11
/
the length scale associated with this vorticity jump caused by the vortices is then equal to
=
_

(u
/L
11
)
= L
11
Re
1/2
0
. (6.2)
In a situation such as the ow generated in an oscillating grid (Holzner et al. (2007)) no mean shear
exists and therefore the coherent vortices should be similar to the worms observed in simulations of
isotropic turbulence. For these structures the strain rate acting on the vortices comes mainly from the
background turbulence eld and is equal to S u
/. Therefore the associated length scale of the

vorticity jump i.e. the scale of the vortices near the T/NT interface is equal to
=
_

(u
/)
= Re
1/2
, (6.3)
where Re
= u
/ is the Reynolds number based on the Taylor micro-scale. Thus the problem of
determining the scale of the T/NT interface becomes the problem of estimating the size of the larger
coherent vortices existing near the edge of the T/NT interface.
An alternative solution can be formulated if one writes from the outset that the vorticity jump observed
at the T/NT interface is caused exclusively by the LVS from the ow:
R
lsvs
. Since for long lived
vortices the timescale associated the the radial viscous diffusion of vorticity is roughly balanced by the
axial stretching caused by the local strain rate eld S, acting on the vortices - as in a Burgers vortex
- the radius of the LVS has to be of the order of the Burgers radius R
lsvs
R
B
=
_
S
_
1/2
. In a jet
the magnitude of the strain rate acting on a LVS near the jet edge is S u
/L
11
leading to R
lsvs

_

(u
/L
11
)
= L
11
Re
1/2
0
, whereas in a very high Reynolds number jet or in a shear free ow, where
the only visible structures are IVS (since the LVS are non existent or are too much fragmented) the strain
is imposed only by the background turbulence and is S u
/ whereby R
ivs

_

(u
/)
= Re
1/2
.
6.3.2 The orientation of the coherent vortices with respect to the T/NT interface
In this section the study of the IVS orientation in relation to the surface dening the T/NT interface
is presented. The IVS are a particular set of vortices taken from the sea of eddies of different sizes,
but since often the LVS were seen to have IVS at their core this orientation reects to some extent
the orientation of the LVS. A similar study, but focusing on the whole vorticity eld, as opposed to the
vorticity contained on the axis of the LVS was carried out by Holzner et al. (2009).
The orientation of the IVS in respect to the T/NT interface is studied through the conditional statistics
and the probability density function (PDF) of cos(
0
) =
0
. n
int
/([
0
[.[ n
int
[), where
0
is the axial vorticity
vector for an identied IVS and n
int
is the local tangent to T/NT interface in the (x, y) plane containing
this IVS axis. The conditional mean prole of cos(
0
) is shown in Fig. 6.5 (a). Figure 6.5 (b) shows
the probability density function (PDF) of cos(
0
) using data from three sets of data, each containing
information from three intervals (distances) from the T/NT interface: [5, 10], [30, 40], and [90, 100].
Very close to the T/NT interface at y
I
/ = 5 10 there is cos(
0
)
I
0.8 (Fig. 6.5 a), which shows
that the IVS are strongly aligned with the local tangent to the T/NT interface. The PDFs of cos(
0
)
conrm the existence of a strong tendency for an alignment between the IVS and the tangent to the
T/NT interface and show that this trend weakens as the centre of the jet shear layer is approached.
These results are not surprising: since the IVS cannot exist outside the turbulent region and cannot
end inside the uid (Helmholtzs second theorem) the only vorticity components that can exist very close
to the T/NT interface are either
x
or
z
but not
y
. The fact that cos(
0
) > 0.5 inside the turbulent region
can be explained by the existence of some remaining legs from streamwise vortices. These results are
also consistent with the conditional mean vorticity proles described in gure 6.2 (b). Furthermore the
strong tendency for an alignment of the IVS with the tangent to the T/NT interface is consistent with
the anisotropy of the viscous dissipation rate observed by da Silva and Pereira (2008): the presence
of dominating axial and azimuthal vortices near the T/NT interface implies that the most intense local
gradient near the interface is u/y, since the axial vortices emerge in opposite pairs, cancelling their
effect on the velocity gradient near the T/NT interface. From this it follows that S
2
xy
> S
2
xz
S
2
yz
as
observed by da Silva and Pereira (2008), which supports the idea that the irrotational viscous dissipa-
tion observed near the T/NT interface is induced by large-scale eddy motions near the jet edge. This
is further supported by the analysis of the LVS near the T/NT interface as described below and agrees
with the recent results reported by Holzner et al. (2009) (see also da Silva et al. (2007)).
<cos(
0
)>
I
0.5
0.6
0.7
0.8
y
I
/
0 25 50 75
(a)
[5,10]
[40,50]
[90,100]
0.1
1
10
cos(
0
)
0 0.5 1
(b)
Figure 6.5: Statistics of cos(
0
) =
0
. n
int
/(|
0
|.| n
int
|), where
0
is the axial vorticity vector for each identied IVS
and n
int
is a local tangent to T/NT interface for the same axial position: (a) conditional proles of cos(
0
); (b) PDFs
of cos(
0
) using data from several intervals from the T/NT interface. The vertical dashed line at y
I
/ = 20 is one
Taylor micro-scale from the T/NT interface.
6.3.3 The kinetic energy viscous dissipation near the T/NT interface
An interesting issue regarding the dynamics of the turbulent quantities near the T/NT interface is the
existence of non-negligible kinetic energy viscous dissipation = 2S
ij
S
ij
, outside the turbulent region
and very close to the T/NT interface as indicated by the experimental results from Holzner et al. (2007,
2008) and the numerical simulations from da Silva and Pereira (2008). The irrotational dissipation
can be seen as anomalous, since no viscous effects can exist in irrotational ow, as described below.
Notice however that in practice, both in the experiments and in the simulations cited above, the vorticity
is different from zero in the irrotational ow because the T/NT interface is dened by points where the
vorticity is greater than a given (although small) vorticity threshold.
It can been argued that no viscous dissipation can exist inside a potential ow region since (Tennekes
and Lumley (1972) p. 77) the Navier-Stokes equations can be written as,
u
i
t
=

x
i
_
p
+
1
2
u
j
u
j
_
+
ijk
u
j
ijk
k
x
j
, (6.4)
where the last term from the right-hand side of the equation is responsible by viscous effects. This
equation shows that viscous effects are absent from irrotational ow since
k
= 0 implying that the
last term from the above equation is zero inside the irrotational ow region. Notice however that the
governing equation for the kinetic energy can be obtained by multiplying u
i
to the above equation which
yields,
_
1
2
u
i
u
i
_
t
=

x
i
_
u
i
p
+
1
2
u
i
u
j
u
j
_
+
ijk
u
i
u
j
ijk
u
i
k
x
j
, (6.5)
where again the effects of molecular viscosity are contained in the last term from the right-hand side
of the equation. Recall that in a turbulent ow the molecular viscosity is always associated with two
effects: viscous diffusion and viscous dissipation. Indeed what the last equation implies is that the sum
of these two effects must be zero in an irrotational ow. Therefore the effect of viscous dissipation
alone is not excluded from the equations. However for viscous dissipation to be active, the net result
of the viscous diffusion has to be positive. At least for the enstrophy a positive net viscous diffusion
contribution has been observed to exist near the T/NT interface inside the irrotational region by da Silva
and Pereira (2007a); Holzner et al. (2008). A possible explanation for this anomalous dissipation has
been suggested by da Silva and Pereira (2008): the analysis of the individual terms contained in S
ij
S
ij
in the plane turbulent jet has showed that the observed viscous dissipation is anisotropic outside the
turbulent region, which is consistent with it being caused by the large scale motions imposed by the LVS
near the T/NT interface (see da Silva and Pereira (2008)). This picture is conrmed here in Figs. 6.6 (a)
and (b). Figure 6.6 (a) shows contours of intense values of viscous dissipation near the T/NT interface
and the LVS of the ow. The visualisations do conrm that intense values of near the T/NT interface
tend to occur whenever nearby LVS are present. Figure 6.6 (b) shows instantaneous proles of and
[ [ across the shear layer. Many similar instantaneous proles for these quantities were observed. The
gure shows a clear link between the vorticity eld associated to a strong LVS near the T/NT interface
and intense values of dissipation originated in the pure shear motion caused by the LVS. Thus the
observed irrotational dissipation originates in the vicinity of the LVS that close to the T/NT interface.
(a)
||
15
30
z/H
0 2 4
0.01
0
(b)
Figure 6.6: (a) Instantaneous contours of vorticity norm | | in a (x, z) plane from the turbulent the jet. The T/NT
interface is indicated by the black line and the values of viscous dissipation outside the turbulent region are
represented by coloured contour lines. The gure shows also the identied IVS (yellow) and the LVS (white) in this
plane. Note that near the T/NT interface there are always signs of the presence of well dened LVS and that
intense viscous dissipation occurs always nearby the most intense LVS; (b) Instantaneous proles vorticity norm
| | across the jet shear layer and viscous dissipation ouside the turbulent region.
<||>
I
<T
4
>
I
1
0
1
2
3
4
y
I
/
0 25 50 75
(a)
dif
(r)/4
0
(r)/
0
-1
-0.5
0
0.5
1
r
0 1 2
(b)
||x10 T
4
100
0
100
200
(c)
Figure 6.7: (a) Conditional proles of enstrophy viscous diffusion T
4
=

2
x
j
x
j
1
2
in relation to the
distance from the T/NT interface; (b) Prole of a model LVS (equation 6.8) and the corresponding values of
enstrophy viscous diffusion

2
x
j
x
j
1
2
(equation 6.9); (c) Instantaneous prole of vorticity norm and

enstrophy viscous diffusion near T/NT interface at the upper shear layer.
6.3.4 The anomalous enstrophy viscous diffusion near the T/NT interface
Another interesting issue concerning viscous effects in the proximity of the T/NT interface concerns
the behaviour of the enstrophy viscous diffusion. The transport equation of this quantity is particularly
interesting to analyse in the context of the turbulent entrainment since the existence or absence of
enstrophy is the most distinguished feature separating the turbulent from the nonturbulent ow regions
(Corrsin and Kistler (1955)). The enstrophy transport equation is given by
_
1
2
i
_
t
+ u
j
_
1
2
i
_
x
j
=
i
j
S
ij
+
2
_
1
2
i
_
x
j
x
j
i
x
j
i
x
j
, (6.6)
where the two terms from the left-hand side of the equation represent the temporal and local variation
of enstrophy and the three terms on the right-hand side represent the enstrophy production, viscous
diffusion, and viscous dissipation, respectively. In most turbulent ows the viscous diffusion term is the
least important regarding their relative magnitudes and in isotropic turbulence it becomes identically
zero. Near the T/NT interface, however, this term exhibits a curious shape. Experimental data from
Holzner et al. (2007, 2008) and numerical simulations by da Silva and Pereira (2007a); da Silva et al.
(2007) show that the enstrophy viscous diffusion term is negligibly small in most of the turbulent and
irrotational regions except very close to the T/NT interface with a negative minimum inside the turbulent
region at y
I
/ 10 and a positive maximum right at the T/NT interface at y
I
/ 0 (see Fig. 6.7 a).
The mechanisms responsible for this curious shape have been analysed in detail in Holzner et al.
(2008) in terms of the relationships between strain and enstrophy. However a more simple explanation
can be given knowing that the T/NT interface location is dened by the LVS. The vorticity radial prole
of a LVS can be well approximated by a Gaussian function,
(r)/
0
= e
r
2
/R
2
(6.7)
so that the enstrophy radial prole is,
(r)
2
/
2
0
= e
2r
2
/R
2
. (6.8)
6.4 Summary 93
Consequently, the radial prole of enstrophy diffusion is,
2
r
2
_
1
2
(r)
2
_
=
4
2
0
R
2
e
2r
2
/R
2
_
1
4r
2
R
2
_
. (6.9)
Figure 6.7 (b) shows the radial proles corresponding to equations (6.8) and (6.9) for the case
0
=
R = 1 for sake of argument. As can be seen, the viscous enstrophy diffusion term has the same shape
observed near the T/NT interface, with negative values inside the vortex and positive values just outside.
This is of course explained by the mechanics of a diffusion term, removing high values of the diffused
quantity where they are above the average (inside the vortex - hence the negative value) and injecting it
into regions where it is lower than the average - hence the positive value outside the vortex, representing
a net production of enstrophy as described in Holzner et al. (2008). Finally, Fig. 6.7 (c) shows an
instantaneous prole of vorticity norm and enstrophy viscous diffusion near T/NT interface at the upper
shear layer. The chosen line goes trough the centre of two big vortices that can be easily identied by the
local peaks of vorticity. As can be seen the enstrophy viscous diffusion near these structures exhibits
a similar structure as predicted by the model enstrophy and enstrophy viscous diffusion in equations
(6.8) and (6.9), with (negative) minima at the centre of the vortices and positive maxima just outside the
vortex cores.
In summary, these results show that the so called anomalous enstrophy diffusion observed near the
T/NT interface correspond to the viscous diffusion process taking place around the LVS sitting near the
T/NT interface. It could be argued that the so called nibbling eddy motions responsible for the turbulent
entrainment process are in effect the viscous diffusion of vorticity from the LVS near the T/NT interface,
but this has to be assessed carefully in a future work.
6.4 Summary
A DNS of a turbulent plane jet at Re
= 120 was used to analyse the role of the coherent vortices near
the T/NT interface. These structures can be divided into large-scale vortex structures (LVS) maintained
by the mean shear and intense vorticity structures (IVS) created by the background turbulence.
The shape of the T/NT interface is dened by the characteristics of these structures since ultimately
it is these structures that dene this interface. The characteristic vorticity jump of the T/NT interface
as well as its thickness is dictated by the radial vorticity distribution of the bigger LVS near the T/NT
interface. Specically the radius of these LVS is equal to the Taylor micro-scale and explains why this
is also the average thickness of the T/NT interface. The coherent vortices in the proximity of the T/NT
interface are preferentially aligned with the tangent to the T/NT interface and are responsible for the
viscous dissipation of kinetic energy observed outside the turbulent region.
Finally it is argued that the nibbling eddy motions responsible for the turbulent entrainment are the
manifestation of the enstrophy viscous diffusion that exists near the edges of the LVS sitting near the
T/NT interface. The lifetime and dynamics of these structures will be the subject of a future study.
The results in the chapter were published in da Silva and dos Reis (2011).
CHAPTER 7
Intense vorticity structures IVS near the turbulent/non-turbulent (T/NT)
interface
This chapter presents the study of the role of IVS in the Turbulent/Non-Turbulent interface and a
possible answer for the presence of dissipation outside of the turbulent zone. First the DNS simulations
and numerical analysis tools described. Then follows the general characteristics of the IVS and LVS in
the whole ow and nally a more detailed analysis near the T/NT interface.
7.1 Direct numerical simulation of a turbulent plane jet
In this chapter the intense vorticity structures are analysed near the turbulent/non-turbulent interface
by using a temporal direct numerical simulation (DNS) of a turbulent plane jet, which will be referred to
by P. Jet. This is the same DNS already presented in the previous chapter.
The work presented uses a vortex tracking procedure to detect and characterise the IVS. In order
to validate this vortex tracking procedure a small DNS of (forced) isotropic turbulence was also carried
out. This simulation is denoted by HIT. This simulation was made using a standard pseudo-spectral
code with low wave number forcing described in da Silva and Pereira (2007b). The simulation uses
256
3
collocation points and the Reynolds number based on the Taylor micro-scale is Re
= 111 and the

maximum resolved wave number multiplied by the Kolmogorov micro-scale is k
max
= 1.51.
A rst comparison of the ow structures present in the two simulation - HIT and P. Jet - can be
made using histograms representing the fraction of vorticity and enstrophy associated with points above
a given threshold. These histograms are displayed in gure 7.1. In HIT the histograms for the three
vorticity components are equal, which is a consequence of the statistical isotropy of the vorticity eld.
The same occurs in the plane jet for vorticity values corresponding to the most frequent events e.g.,
96 Intense vorticity structures IVS near the turbulent/non-turbulent (T/NT) interface
0 2 4 6 8
10
-4
10
-3
10
-2
10
-1
10
0
|
x
|/
|
y
|/
|
z
|/
||/
Plane Jet
0 2 4 6 8
10
-4
10
-3
10
-2
10
-1
10
0
Isotropic
Figure 7.1: One dimensional histograms of the volume fraction occupied by points above a certain vorticity
threshold in isotropic turbulence (HIT) and in the plane jet (P. Jet). The x axis represents e.g. |
x
|/
, with
= (
i
)
1/2
, where
i
is the uctuating vorticity eld. For the jet case
roughly constant inside the shear layer

and a mean value was taken.
[
x
[/
< 1, and also, approximately, for the intense vorticity e.g., [

x
[/
> 1.
7.2 The intense vorticity structures inside the shear layer region
In the present work the statistics of the IVS such as their core radius R and maximum tangential
velocity u
0
were determined using a tracking procedure described in sec. (5.2). In (forced) isotropic tur-
bulence Jimnez et al. (1993); Jimnez and Wray (1998) showed that the vortex core size R, maximum
axial vorticity
0
, circulation Reynolds number Re
= /, and maximum azimuthal velocity u

0
for the
IVS scale as R 5,
0

Re
1/2
, Re
20Re
1/2
, and u
0
u
, respectively, where is the circulation,

and
0
= (
i
j
S
ij
/
k
k
)
0
is the axial stretching rate at the axis of the IVS. It has been shown moreover,
that in forced isotropic turbulence the mean radius of the IVS is R R
B
, where R
B
= 2 (/
0
)
1/2
is
the Burgers radius i.e., the IVS from forced isotropic turbulence are approximately equilibrium Burgers
vortices Jimnez and Wray (1998).
Table 7.1 presents a compilation of the IVS characteristics found in several different ows at several
Reynolds numbers, such as (forced and decaying) homogeneous isotropic turbulence, homogeneous
shear, mixing layer, round and plane jets, circular duct, and channel and boundary layer ows. The
results obtained with the present tracking procedure in the turbulent plane jet and in (forced) isotropic
turbulence (HIT) are also shown.
The values of the mean radius of the vortex core divided by the Kolmogorov micro-scale R / are
close to R / 45 for a variety of turbulent ows and for a wide range of Reynolds numbers. For the
present HIT and P. Jet the result is R / = 4.6 for Re
= 111 and 120, respectively which is well within

the range of values found in the literature. The mean tangential velocity u
0
divided by the root-mean
square velocity u
obtained in the literature varies between 0.50 < u

0
/u
< 1.21 indicating that u

0
u
.
The result is 0.68 and 0.76 for HIT and for P. Jet, respectively, which is again well within the values found
in the literature. The same is true for the mean tangential velocity normalised by the Kolmogorov velocity
7.2 The intense vorticity structures inside the shear layer region 97
scale u
. Finally, the mean axial vorticity of the IVS can be compared by computing the Reynolds number
based on the vortex circulation Re
= /, divided by the square-root of the Reynolds number based

on the Taylor micro-scale Re
1/2
, where the circulation is obtained through =

_
u.
dl = 2Ru
0
. Once
more, the present values of Re
/Re
1/2
= 28.8 (HIT) and 28.3 (P. Jet) are in good agreement with the
values obtained in the literature 10.5 < Re
/Re
1/2
< 32.4. The same can be said of the axial vorticity

which is
0
/
_
Re
1/2
_
= 0.39 in the forced isotropic turbulence conguration and
0
/
_
Re
1/2
_
= 0.38
in the turbulent plane jet, and that compares well with 0.30 <
0
/
_
Re
1/2
_
< 0.42 observed in isotropic
turbulence in Jimnez and Wray (1998). These results show that the tracking procedure used here is
indeed well implemented.
F
l
o
w
S
o
u
r
c
e
R
e
_
R
R
B
_
u
0
u
0
R
e
R
e
1
/
2
F
.
I
s
o
t
r
o
p
i
c
:
J
i
m
n
e
z
e
t
a
l
.
(
1
9
9
3
)
3
6
-
1
6
8
3
.
8
-
4
.
2
1
6
.
5
-
2
1
.
1
J
i
m
n
e
z
a
n
d
W
r
a
y
(
1
9
9
8
)
3
7
-
1
6
8
4
.
8
-
4
.
9
0
.
9
2
-
1
.
0
5
0
.
8
6
-
0
.
9
9
K
i
d
a
a
n
d
M
i
u
r
a
(
1
9
9
8
)
4
6
3
.
2
-
3
.
8
1
4
.
7
-
3
2
.
4
p
r
e
s
e
n
t
H
I
T
1
1
1
4
.
6
0
.
9
9
0
.
6
8
9
.
0
2
8
.
8
D
.
I
s
o
t
r
o
p
i
c
:
J
i
m
n
e
z
a
n
d
W
r
a
y
(
1
9
9
8
)
6
2
4
.
8
0
.
8
2
1
.
2
1
H
.
S
h
e
a
r
:
K
i
d
a
a
n
d
M
i
u
r
a
(
1
9
9
8
)
4
.
9
-
5
.
2
M
.
L
a
y
e
r
:
T
a
n
a
h
a
s
h
i
e
t
a
l
.
(
2
0
0
1
)
8
0
-
1
0
0
4
.
5
0
.
5
0
D
u
c
t
:
M
o
u
r
i
e
t
a
l
.
(
2
0
0
7
)
7
1
9
-
1
9
3
4
5
.
5
-
6
.
2
1
0
.
5
9
-
0
.
6
4
8
.
7
-
1
4
.
0
1
0
.
5
-
1
2
.
3
B
.
l
a
y
e
r
:
M
o
u
r
i
e
t
a
l
.
(
2
0
0
7
)
3
3
2
-
1
3
0
4
5
.
2
-
6
.
2
1
0
.
6
8
-
0
.
8
2
7
.
6
-
1
2
.
8
1
3
.
6
-
1
3
.
7
C
h
a
n
n
e
l
:
K
a
n
g
e
t
a
l
.
(
2
0
0
8
)
2
0
0
-
3
8
0
4
.
0
-
5
.
0
1
.
2
-
2
.
0
R
.
j
e
t
:
G
a
n
a
p
a
t
h
i
s
u
b
r
a
m
a
n
i
(
2
0
0
8
)
1
5
0
3
.
0
-
7
.
5
P
.
j
e
t
:
d
a
S
i
l
v
a
a
n
d
P
e
r
e
i
r
a
(
2
0
0
8
)
1
2
0
4
.
6
0
.
9
7
0
.
7
6
7
.
1
5
2
8
.
3
T
a
b
l
e
7
.
1
:
C
h
a
r
a
c
t
e
r
i
s
t
i
c
s
o
f
t
h
e
I
V
S
i
n
d
i
f
f
e
r
e
n
t
o
w
s
a
n
d
R
e
:
f
o
r
c
e
d
i
s
o
t
r
o
p
i
c
t
u
r
b
u
l
e
n
c
e
(
F
.
I
s
o
t
r
o
p
i
c
)
,
d
e
c
a
y
i
n
g
i
s
o
t
r
o
p
i
c
t
u
r
b
u
l
e
n
c
e
(
D
.
I
s
o
t
r
o
p
i
c
)
,
h
o
m
o
g
e
n
e
o
u
s
s
h
e
a
r
(
H
.
S
h
e
a
r
)
,
m
i
x
i
n
g
l
a
y
e
r
(
M
.
L
a
y
e
r
)
,
r
o
u
n
d
a
n
d
p
l
a
n
e
j
e
t
s
,
(
R
.
j
e
t
,
P
.
j
e
t
)
c
i
r
c
u
l
a
r
d
u
c
t
(
D
u
c
t
)
,
c
h
a
n
n
e
l
a
n
d
b
o
u
n
d
a
r
y
l
a
y
e
r
o
w
s
(
C
h
a
n
n
e
l
,
B
.
L
a
y
e
r
)
.
M
e
a
n
v
a
l
u
e
s
o
f
r
a
d
i
u
s
R
,
a
z
i
m
u
t
h
a
l
v
e
l
o
c
i
t
y
u
0
,
a
n
d
R
e
y
n
o
l
d
s
n
u
m
b
e
r
b
a
s
e
d
o
n
t
h
e
c
i
r
c
u
l
a
t
i
o
n
R
e
f
o
r
t
h
e
I
V
S
,
n
o
n
-
d
i
m
e
n
s
i
o
n
a
l
i
z
e
d
b
y
t
h
e
R
e
y
n
o
l
d
s
n
u
m
b
e
r
b
a
s
e
d
o
n
t
h
e
T
a
y
l
o
r
m
i
c
r
o
-
s
c
a
l
e
R
e
,
K
o
l
m
o
g
o
r
o
v
m
i
c
r
o
-
s
c
a
l
e
,
B
u
r
g
e
r
s
v
o
r
t
e
x
r
a
d
i
u
s
R
B
,
K
o
l
m
o
g
o
r
o
v
v
e
l
o
c
i
t
y
s
c
a
l
e
u
,
a
n
d
r
o
o
t
-
m
e
a
n
s
q
u
a
r
e
v
e
l
o
c
i
t
y
u
.
T
h
e
R
e
y
n
o
l
d
s
n
u
m
b
e
r
s
i
n
K
a
n
g
e
t
a
l
.
(
2
0
0
8
)
w
e
r
e
e
s
t
i
m
a
t
e
d
f
r
o
m
t
h
e
d
a
t
a
p
r
e
s
e
n
t
e
d
i
n
t
h
e
p
a
p
e
r
.
N
o
t
i
c
e
t
h
a
t
t
h
e
d
e
n
i
t
i
o
n
f
o
r
R
e
u
s
e
d
h
e
r
e
i
s
s
l
i
g
h
t
l
y
d
i
f
f
e
r
e
n
t
t
o
t
h
e
d
e
n
i
t
i
o
n
u
s
e
d
b
y
s
o
m
e
a
u
t
h
o
r
s
w
h
e
r
e
R
e
/
(
2
)
.
I
n
t
h
e
s
e
c
a
s
e
s
t
h
e
v
a
l
u
e
s
f
o
u
n
d
i
n
t
h
e
l
i
t
e
r
a
t
u
r
e
w
e
r
e
c
o
n
v
e
r
t
e
d
t
o
t
h
e
d
e
n
i
t
i
o
n
o
f
R
e
u
s
e
d
h
e
r
e
.
F
o
r
t
h
e
p
l
a
n
e
j
e
t
c
o
n
g
u
r
a
t
i
o
n
t
h
e
r
e
f
e
r
e
n
c
e
u
,
a
n
d
a
r
e
t
a
k
e
n
f
r
o
m
t
h
e
c
o
n
d
i
t
i
o
n
a
l
m
e
a
n
p
r
o
l
e
s
o
f
t
h
e
s
e
q
u
a
n
t
i
t
i
e
s
,
a
n
d
e
a
c
h
i
n
s
t
a
n
t
a
n
e
o
u
s
e
l
d
u
s
e
s
i
t
s
c
o
r
r
e
s
p
o
n
d
i
n
g
r
e
f
e
r
e
n
c
e
v
a
l
u
e
s
f
o
r
,
u
a
n
d
w
h
i
c
h
a
r
e
r
o
u
g
h
l
y
c
o
n
s
t
a
n
t
d
e
e
p
i
n
s
i
d
e
t
h
e
t
u
r
b
u
l
e
n
t
r
e
g
i
o
n
a
n
d
s
u
b
j
e
c
t
e
d
t
o
s
m
a
l
l
c
h
a
n
g
e
s
f
o
r
t
h
e
N
T
=
1
1
e
l
d
s
u
s
e
d
h
e
r
e
.
7.2 The intense vorticity structures inside the shear layer region 99
JET
HIT
10
3
10
2
10
1
10
0
R/
0 5 10
(a)
JET
HIT
10
3
10
2
10
1
10
0
u
0
/u'
0 1 2
(b)
JET
HIT
0
0.02
0.04
Re
/Re
1/2
0 50 100
(c)
JET
HIT
10
-3
10
-2
10
-1
10
0
R/R
B
0 1 2 3 4
(d)
Figure 7.2: Pdfs of the lament properties in HIT and in the P. Jet (data from the entire shear layer region): (a)
radius normalised with the Kolmogorov micro-scale; (b) maximum tangential velocity normalised with the rms
velocity; (c) circulation Reynolds number normalised with the Taylor-scale Reynolds number; (d) radius normalised
with the Burgers radius. For the P. Jet the values used in the normalisation e.g., u
are the values deep inside the

shear layer and each instantaneous eld is normalised by its corresponding value of u
.
In order to compare the characteristics of the IVS in isotropic turbulence with the IVS in the tur-
bulent plane jet gs. 7.2 (a-d) show the probability density functions (pdfs) of the vortex radius non-
dimensionalized by the Kolmogorov micro-scale , the tangential velocity non-dimensionalized by the
root-mean square velocity u
, the circulation Reynolds number normalised with the Reynolds number,

and the vortex core radius normalised with the Burgers radius, respectively.
The shape and the magnitudes of the pdfs shown in Figs. 7.2 (a-d) agrees remarkably well with the
same pdfs displayed in e.g., Jimnez et al. (1993); Jimnez and Wray (1998) and in Tanahashi et al.
(2001). Notice however that tails of the pdfs in the plane jet show some differences compared to isotropic
turbulence e.g., the probability of nding vortex core radius with R = 10 is slightly higher in a jet than in
isotropic turbulence (Fig. 7.2 a). Also, the extreme values of the tangential velocities found in the jet are
also slightly more frequent than in isotropic turbulence (Fig. 7.2 b).
The differences between the IVS found in a jet and in isotropic turbulence will be analysed below,
however it is interesting to notice that the pdf of R/R
B
is very similar in a jet and in isotropic turbulence
which indicates that the IVS found in the jet are well described by the Burgers vortex model. The Burgers
vortex model occupies an important place in the study of vortical structures in turbulent ows since it
is known from Jimnez and Wray (1998) that the worms from forced isotropic turbulence are very well
described by this model.
The steady Burgers vortex is an exact solution of the Navier-Stokes equations describing a vortex
tube immersed in an axisymmetric, irrotational eld (Davidson (2004)). In this model vortex the axial
vorticity is given by
z
(r) =

4
e
r
2
/R
2
B
(7.1)
and the velocity eld in cylindrical coordinates is
u
z
= z u
r
=
1
2
r u
=

2r
_
1 e
r
2
/R
2
B
_
(7.2)
where = 2
_
0

z
(r)rdr is the vortex circulation, is the rate-of-strain and R
B
= 2(/)
1/2
is the
radius of the Burgers vortex, which is kept constant due to the balance between the axial stretching rate
and the radial viscous diffusion. Jimnez and Wray (1998) showed that the axial stretching in forced
isotropic turbulence originates in the background vorticity consisting of structures for which
< <
ivs
. In this ow the Burgers vortex radius is R
B
= 2(/
0
)
1/2
where
0
=
0
T
.S.
0
/[
0
[
2
is the axial
stretching rate acting on the axis of each IVS (with axial vorticity
0
) and S is the local rate-of-strain. The
ratio between the vortex core radius and the Burgers radius R/R
B
in table 7.1 shows that in (forced)
isotropic turbulence R/R
B
= 0.99 which is very close to the values obtained by Jimnez and Wray
(1998). For the plane jet the result is R/R
B
= 0.97 which is slightly less than in isotropic turbulence but
somehow also supports the idea that the IVS from the planar jet are, generally i.e. in the whole shear
layer, well described by the equilibrium Burgers vortex model.
7.3 The IVS near the T/NT interface
7.3.1 Topology of the IVS near the T/NT interface
Turning into the analysis of the IVS near the T/NT interface. It is instructive to have a glimpse of
these structures near the T/NT interface. Before being able to do this, the exact location of the interface
separating turbulent and irrotational ow has to be dened. Since the T/NT interface divides the ow
into a rotational (turbulent) and an irrotational (non-turbulent) region, the vorticity (or the enstrophy) is
the appropriate variable to dene the exact location of the interface. Following several previous works
(e.g., Bisset et al. (2002); Mathew and Basu (2002); da Silva and Pereira (2008); da Silva (2009) the
T/NT interface location was dened as the ow surface where the local vorticity norm = (
i
i
)
1/2
is
equal to a certain threshold. Recently Anand et al. (2009) compared several different criteria to dene
the T/NT interface.
Figures 7.3 (a-d) and 7.4 (a-d) show the ow structures near the T/NT interface in the upper shear
layer of the plane jet allowing to understand the geometry and interplay of the several structures. Figure
7.3 (a) displays the IVS (yellow iso-surfaces) in sections of small circular disks, where the radius of each
disk is the exact radius of the local IVS radius computed with the vortex tracking algorithmused here. The
length and radial dimension of the IVS is not uniform but the majority of the structures seem to display
similar characteristics. The IVS do not show a clear spatial orientation, like in isotropic turbulence, but
seemto be clustered in two separate regions, at the borders of the ow domain. Figure 7.3 (b) shows iso-
surfaces of intense enstrophy (blue) corresponding to the background vorticity i.e., structures for which
7.3 The IVS near the T/NT interface 101
(a) (b)
(c) (d)
Figure 7.3: Side view of the coherent structures near the T/NT interface in the upper shear layer of the turbulent
plane jet: (a) IVS detected by the vortex tracking algorithm (yellow). Notice that they are displayed to scale. (b)
Iso-surfaces of enstrophy (blue) and IVS (yellow); (c) Iso-surfaces of low pressure (orange) and IVS (yellow); (d)
Iso-surfaces of low pressure regions (orange), strong enstrophy (blue), and IVS (yellow).
(a) (b)
(c) (d)
Figure 7.4: Side view of the coherent structures near the T/NT interface in the upper shear layer of the turbulent
plane jet: (a) Iso-surfaces of the IVS (yellow), low pressure (white), and the T/NT interface (brown); (b) Same as in
(a) but using a transparent iso-surface for the pressure; (c) Same as in (b) but using a transparent iso-surface for
the T/NT interface; (d) Iso-surfaces of low pressure (white), and the T/NT interface (brown).
the enstrophy is
< <
ivs
and IVS (yellow). As can be seen the IVS merge with the background
vorticity indicating that the IVS are extreme events of the background vorticity, like in isotropic turbulence
as described in Jimnez et al. (1993). The longer enstrophy structures seem to consist of pairs of
vortices aligned preferentially with the streamwise direction. Figure 7.3 (c) shows iso-surfaces of low
pressure (orange), and IVS (yellow). The pressure eld allows one to observe the large scale structures
of the ow showing three large scale structures, aligned with the z direction: two at the upper shear
layer, at the extremes of the domain, and another one below, at the lower shear layer. A considerable
amount of streamwise structures is also observed. The radius of the spanwise rollers R
lvs
was estimated
in da Silva and Taveira (2010) and is close to the Taylor micro-scale R
lvs
. The structures are
remnants of the Kelvin-Helmholtz rollers that are known to arise during the early stages of the transition
to turbulence in a jet. Comparing Figs. 7.3 (a) and (c) one sees that the clusters of IVS described
before are located inside the large scale vortices identied by the low pressure iso-surfaces, however
the correlation between the large scale structures and the IVS is apparently low, which underscores the
scale separation between the two types of structures. Finally, gure 7.3 (d) shows all three iso-surfaces
i.e., IVS (yellow), enstrophy (blue), and pressure (orange) and shows that there is some overlap between
the enstrophy structures and the pressure but only for the bigger enstrophy structures, such as one of
the two streamwise vortices which seem to be merging into a single structure (compare also Figs. 7.3
b and c) i.e. the spatial overlap between the LVS and the IVS (low pressure is associated with strong
vorticity), is clearly apparent from the gure, but one can see again that the IVS exhibit more "random"
spatial orientation than the LVS.
In order to analyse the interplay between these structures and the T/NT interface gures 7.4 (a-d)
use partially transparent (or translucent) iso-surfaces of the same quantities displayed in gures 7.3
(a-d) and show also the T/NT interface at the upper shear layer of the jet. Following several previous
works e.g., Bisset et al. (2002); Mathew and Basu (2002); da Silva (2009) the T/NT interface consists
on the surface dened by a certain vorticity norm threshold = (
i
i
)
1/2
, where
i
= u
i
is the
vorticity eld and the detection threshold is = 0.7U
1
/H, which is the same threshold used in Bisset
et al. (2002); Mathew and Basu (2002). Fig. 7.4 (a) shows part of the T/NT interface (brown) and the
LVS (white) identied using pressure iso-surfaces, as well as some IVS (yellow). The T/NT interface is
strongly convoluted, something that is well known for a long time, and one sees immediately that the
length scale of the convolutions is very similar to the length scale of the LVS below its surface. Figures
7.4 (b) and (c) are equivalent to Fig. 7.4 (a) but show the pressure iso-surface as translucent (Fig. 7.4 b)
and the T/NT interface as translucent (Fig. 7.4 c) which allows one to observe the IVS below the T/NT
interface. Contrary to the LVS, the IVS play no role on the denition of the T/NT interface characteristics,
since they are neither of the same scale of the observed T/NT interface convolutions nor are they the
rst structures that appear right below the T/NT interface. With Fig. 7.4 (d) displaying the full T/NT
interface at the upper shear layer one realises, not surprisingly, that the observed convolutions of the
T/NT interface are dictated by the large vorticity structures (LVS) underneath its surface. Indeed the
T/NT interface coincides with the physical line dened by the LVS at the edge of the upper shear layer
and this fact underlines one of the well know features of the T/NT interface that has been known for
some time (e.g. Townsend (1976)): the LVS are responsible for the observed convolutions of the T/NT
interface and this explains why the length scale of these convolutions is roughly equal to the length scale
of the LVS.
7.3.2 Conditional statistics of the IVS characteristics near the T/NT interface
In order to study the inuence of the T/NT interface upon the IVS characteristics conditional statistics,
in relation to the distance from the T/NT interface, were used. The procedure used to obtain these
statistics has been used in several previous works (e.g., Bisset et al. (2002); Westerweel et al. (2005,
2009); da Silva and Pereira (2008); da Silva (2009)) and is described in chapter 5.
Notice that here as in da Silva (2009) the local coordinate system is always aligned with the y di-
rection, unlike in da Silva and Pereira (2008) where the conditional statistics were made along a line
which is always locally normal to the T/NT interface. It was observed that both procedures lead to very
similar results, however. The same procedure is used also for the lower shear layer and the statistics
use each one of the N
z
spanwise planes and N
T
= 11 instantaneous elds taken from the fully devel-
oped turbulent regime between T/T
ref
= 20.2 to 27.0. Note that the conditional proles shown below
highlight the region in the proximity of T/NT interface, between 0 < y
I
/ < 75, since the variables anal-
ysed here are roughly constant from y
I
/ 75 until the centre of the jet shear layer, which is located
approximately at y
I
/ 140 150. Some of the conditional mean proles presented in this work are
scaled by reference turbulent quantities e.g., R
I
/. Although in general these quantities e.g., or
u
vary in time and along the y direction, their conditional mean proles roughly collapse deep inside
the turbulent region for the N
T
= 11 elds used here. Nevertheless the reference turbulent quantities
used for each eld correspond to their value deep inside the turbulent region e.g., for each eld/in-
stant is taken as =
I
for y
I
/ = 75. The conditional mean proles result from averaging all
the instantaneous proles, where each one was scaled by its "instantaneous" reference quantity i.e.,
R/
I
= R(y
I
)/
I
=
1
N
T
N
T
=11
i=1
R(y
I
)
I
/
i
. Since the turbulence elds are from the self-similar
turbulent regime it is assumed that the turbulent characteristics studied here are representative of the
ow in turbulent jets.
Figure 7.5 (a) repeats here, from the previous chapter and for context, the conditional mean proles
of [
x
[
I
, [
y
[
I
and [
z
[
I
in relation to the distance from the T/NT interface, showing a sharp jump
across the T/NT interface with thickness roughly equal to 20. In the present simulation and deep inside
the turbulent region the result is 20, where is the Taylor micro-scale, therefore this jump is
equal to the Taylor micro-scale, in agreement with other experimental and numerical works (Bisset et al.
(2002); Westerweel et al. (2005)). As described in the introduction and as shown by da Silva and Taveira
(2010) this is because the thickness of the T/NT interface is equal to the radius of the LVS near the T/NT
interface
R
lvs
.
Figure 7.5 (b) shows the percentage of points which were identied as belonging to an existing IVS
axis in the region between 0 < y
I
/ < 75 i.e., fraction of the total number of detected IVS in this region,
as function of the distance from the T/NT interface (integrating this prole between 0 < y
I
/ < 75 the
result is 100%). For y
I
/ > 20 25 this number is roughly constant, but it decreases sharply as one
<||>
I
<|
x
|>
I
<|
y
|>
I
<|
z
|>
I
0
1
2
3
4
y
I
/
0 25 50 75
(a)
<||>
I
% axis pts
0
1
2
3
4
y
I
/
0 25 50 75
(b)
Figure 7.5: (a) Mean conditional proles of ||
I
, |
x
|
I
, |
y
|
I
and |
z
|
I
1
/H); (b)
Percentage of the total number of points identied as IVS, as a function of the distance from the T/NT interface in
the region 0 < y
I
/ < 75 i.e., integrating this prole between 0 < y
I
/ < 75 there is 100% (||
I
is also shown).
The vertical dashed line at y
I
/ = 20 is one Taylor micro-scale from the T/NT interface.
approaches the T/NT interface until, for 0 < y
I
/ < 5, this number is negligibly small. This shows, as
expected, that there is a thin region near the T/NT interface where no axis from IVS exist, since this is
roughly the radius of the existing IVS. This agrees also with the results from da Silva and Taveira (2010)
and da Silva and dos Reis (2011): the T/NT interface in fact consists of (or is made up from) LVS sitting
at that location. Finally, notice that (Figure 7.5b) the size of the adjusting region 5 < y
I
/ < 25 where the
IVS are not as numerous as inside the jet is consistent with the analysis of the invariant maps described
in da Silva and Pereira (2008, 2009): at y
I
/ = 20 there are already signs of the presence of LVS.
7.3.3 IVS characteristics near the T/NT interface: the Burgers vortex model
This section analyses the IVS characteristics as function of the distance from the T/NT interface.
Figures 7.6 (a-d) show conditional mean proles of the vortex core radius R (Fig. 7.6 a), axial vorticity
0
(Fig. 7.6 b), azimuthal velocity u
0
(Fig. 7.6 c), and circulation Reynolds number (Fig. 7.6 d). The
conditional mean prole of the vorticity norm [[ for the whole ow eld is also shown (Fig. 7.6 b). The
vertical dashed line at y
I
/ = 20 marks the distance from the T/NT interface which is equal to the Taylor
micro-scale.
The core radius deep inside the jet shear layer (y
I
/ > 75) is equal to R
I
4.3 which is close
to the value observed in isotropic turbulence. However, it increases near the T/NT interface reaching
R
I
5.3 at y
I
/ 20, before decreasing to R
I
4.2 at y
I
/ = 10. It is interesting to see that
the maximum observed radius is located at y
I
/ 20 which is exactly one Taylor scale from the T/NT
interface and is also the point of maximum vorticity as described in da Silva and Pereira (2008). This can
be explained by the dynamics of the IVS near the jet edge as will be described below. The quick drop
of R
I
/ in the region 10 < y
I
/ < 20 has to be caused by the lack of space inside this layer. There
is simply no space for IVS - whose diameter is close to 10 - inside this region. Notice that, as seen
before, the number detected IVS between 0 < y
I
/ < 20 drops fast as the T/NT interface is approached
<R>
I
/
3
4
5
6
y
I
/
0 25 50 75
(a)
<||>
I
<|
0
|>
I
2
3
4
5
y
I
/
0 25 50 75
(b)

I
/u'
0.6
0.7
0.8
0.9
y
I
/
0 25 50 75
(c)
<Re
>
I
/Re
1/2
20
30
40
50
y
I
/
0 25 50 75
(d)
Figure 7.6: Conditional proles of IVS characteristics: (a) vortex core radius R
I
/; (b) axial vorticity
0
(U
1
/H)
(the vorticity norm from the whole eld || is also shown); (c) tangential velocity u
0
I
/u
; (d) circulation Reynolds

number Re
I
/Re
1/2
. The vertical dashed line at y

I
and therefore the statistics from this region have to be considered carefully even though the observed
trend is clear.
Interestingly the axial vorticity of the detected IVS is sensibly constant throughout the jet shear layer
(Fig. 7.6 b), so that even the IVS very close to the T/NT interface have similar axial vorticity to those
found deep inside the shear layer. This indicates that it is mostly the orientation of the IVS and not their
vorticity magnitude that is affected by the presence of the T/NT interface, in agreement with da Silva and
dos Reis (2011) who showed that the IVS close to the T/NT interface tend to be aligned with the tangent
to the interface surface.
The evolution of the conditional azimuthal velocity of the IVS (Fig. 7.6 c) resembles the evolution
of R/ with a maximum close to the T/NT interface u
0
I
/u
0.84 (at y
I
/ = 20) and smaller values
deep inside the shear layer u
0
I
/u
0.72 for y
I
/ > 75. Again, very close to the T/NT interface
10 < y
I
/ < 20 the azimuthal velocity of the IVS decreases very fast, which indicates that the vortex
lament characteristics are modied by the presence of the T/NT interface.
Although the IVS are a particular subset of structures taken from a sea of different eddy sizes it is
already possible to draw one conclusion from the above results: since the core radius and azimuthal
velocity are higher near the T/NT interface (y
I
/ 20) than in the region 10 < y
I
/ < 20, and than deep
inside the turbulent region (y
I
/ > 75), while the axial vorticity is roughly constant across the shear
layer, this implies that the circulation
0
of the IVS is also maximum at y
I
/ 20 has lower values
between 10 < y
I
/ < 20 or in the interior of the shear layer y
I
/ > 20, because the circulation of each
particular IVS is
0
2u
0
R. This is indeed what is observed in the circulation Reynolds number whose
conditional prole is shown in Fig. 7.6 (d).
It is tempting to infer that the observed differences in the circulation of the IVS across the shear layer
is caused by the characteristics of the ow eld near the T/NT interface since the results suggest that
the IVS that are to be found at this location were formed at that particular location and are not structures
that were formed in other parts of the ow e.g. the jet centreline, and that were simply advected into
the jet edge. Although only a detailed analysis of the IVS in time would allow to settle denitively these
questions it is easy to show that the lifetime of the IVS, T
, is much smaller than the time needed for

these structures to travel from the region near the jet centreline into the region near the jet edge T
V
.
As will be shown below the characteristic time associated with the downstream convection of the IVS,
T
U
, is much bigger than the lifetime of these structures T
T
U
, and since the downstream transport,
associated with the higher mean velocity component of the jet, is much smaller than the time associated
with lateral transport of the IVS, it follows naturally that the IVS are dissipated much faster than they can
be transported from the centre of the jet into the region near the jet edge i.e., T
T
U
T
V
.
<T
>
I
<T
S
>
I
<R/R
B
>
<(T
/T
S
)
1/2
>
I
0.5
1
1.5
2
y
I
/
0 25 50 75
Figure 7.7: Conditional proles of diffusion and stretching time-scales T
I
and T
S
I
, and ratio between the local
vortex core radius and the Burgers radius R/R
B
I
I
/ = 20 is one Taylor
micro-scale from the T/NT interface.
The dynamics of the IVS can be analysed by comparing their characteristics with the Burgers vortex
model. As described above it is well known that the Burgers vortex model provides an excellent descrip-
tion of the IVS found in (forced) isotropic turbulence and it is expected that this model will provide a good
description for the IVS in free shear ows e.g., jets, since the IVS are similar in many different ows. The
mean value of R/R
B
for the entire shear layer supports this view with R/R
B
= 0.97 (see table 7.1). In
forced isotropic turbulence Jimnez and Wray (1998) observed that the radius of the IVS is equal to the
Burgers radius R = R
B
for a wide range of Reynolds numbers: 0.96 < R/R
B
< 1.05 for 62 < Re
< 168.
However in decaying isotropic turbulence Jimnez and Wray (1998) observed R/R
B
= 0.82 for Re
= 62
showing that the Burgers vortex is a less accurate model to describe the dynamics of the IVS in decay-
ing turbulence. The jet ow studied here is also in some sense a decaying ow since the rms velocity
decreases slowly with the axial direction (or with time in a temporal simulation) in the far eld self-similar
region. However the presence of a mean shear producing turbulent kinetic energy and thus some addi-
tional local strain (compared to decaying isotropic turbulence) is expected to counter this decay to some
extent. Therefore, in a jet expected to be somehow between the classical cases of forced and decaying
isotropic turbulence.
The analysis of the IVS dynamics was started by comparing the measured vortex core radius R of
each identied structure with the local Burgers radius R
B
= 2 (/
0
)
1/2
through the ratio R/R
B
displayed
in Fig. 7.7. Notice that R
B
can only be computed for axial IVS points where vortex stretching dominates
over compression
0
> 0. The conditional prole of R/R
B
I
is slightly smaller than 1.0 close to the
T/NT interface (R/R
B
I
0.9 between 5 < y
I
/ < 30), attains a maximum of around R/R
B
I
1.0
between 30 < y
I
/ < 50, and decreases to R/R
B
I
0.97 in the centre of the shear layer. Thus, R/R
B
in the jet is indeed between the values observed in forced and in decaying isotropic turbulence. This
shows that the Burgers vortex model, although being still an approximately good description for most of
the IVS in the jet is less accurate in modelling the IVS near the T/NT interface, and that for these IVS
(near the T/NT interface) enstrophy generation by axial vortex stretching and radial enstrophy viscous
diffusion are not in perfect equilibrium, and the vortex core radius of these structures is not stable i.e., it
is evolving in time.
To investigate where this imbalance comes from two timescales associated with the two mechanisms
were dened: T
S
= 1/
0
and T
= R
2
/ (4) associated with axial vortex stretching and radial viscous
diffusion, respectively. If the two timescales are equal T
S
= T
, the vortex core radius is equal to the

Burgers radius R = R
B
= 2 (/
0
)
1/2
, and the vortex core is stable i.e. it does not change in time. If,
on the contrary, the two timescales are not equal the mechanism associated with the smaller timescale
dominates e.g., a vortex where T
< T
S
is subjected to more intense viscous diffusion than axial stretch-
ing, and this will tend to increase its radius in time. The ratio of the conditional timescales
(T
/T
S
)
1/2
_
I
displayed in gure 7.7 shows that this is the dynamical situation for the IVS near the T/NT interface, and
to a smaller extent to many other IVS to the interior of the shear layer, since T
/T
S
I
< 1.0. Notice that
for given
0
and the two time-scales are related to the vortex core radius by (T
/T
S
)
1/2
= R/R
B
, and
the present results comply with this relationship (see Fig. 7.7).
Figures 7.8 (a) to (d) show the instantaneous conditional IVS characteristics for two different times
T/T
ref
= T
22 and T
24 (the mean value resulting from the averaging procedure using N

T
= 11
instantaneous elds discussed in Fig. 7.6 is also shown). This allows one to observe the temporal
evolution of several IVS characteristics during a short time interval. The degree of convergence of these
curves is not perfect due to the small number of samples, but is sufcient to observe the main trends.
One observes that as time progresses from T
22 to T
24 the axial vorticity (Fig. 7.8 b) and the

tangential velocity (Fig. 7.8 c) decrease in the majority of the ow points, while the magnitude of the
vortex core radius increases (Fig. 7.8 b). This is exactly what one would expect in an isolated vortex if
the viscous diffusion of enstrophy is more important than the axial enstrophy stretching. Notice also that
the circulation Reynolds number decreases also (Fig. 7.8 d). As remarked by a referee, this last result
underscores the fact that the dynamics of the IVS is not inviscid, and the circulation of the individual IVS
is not constant, but decreasing in time. This is a natural consequence of the fact that the length scale of
<R>
I
/
<T
*
>
T
*
=21.85
T
*
=24.20
3
4
5
6
y
I
/
0 25 50 75
(a)
<<|
0
|/'>
I
|T
*
>
<|
0
|>
I
/' |T
*
=21.85
<|
0
|>
I
/' |T
*
=24.20
2
3
4
5
y
I
/
0 25 50 75
(b)

I
/u'
<T
*
>
T
*
=21.85
T
*
=24.20
0.5
1
y
I
/
0 25 50 75
(c)
<Re
>
I
/Re
1/2
<T
*
>
T
*
=21.85
T
*
=24.20
0
25
50
y
I
/
0 25 50 75
(d)
Figure 7.8: Temporal evolution of the conditional proles of IVS characteristics, showing the values at T/T
22
and T
24, and the reference mean values (shown in Figs. 7.6): (a) vortex core radius R
I
/; (b) axial vorticity
I
; (c) tangential velocity u
0
I
/u
; (d) circulation Reynolds number Re
I
/Re
1/2
. The vertical dashed line at

y
I
the IVS is of the order of the Kolmogorov micro-scale (R 45) implying that the viscosity is important
for these structures.
The series of observations described above raise a new question: why is the dynamics of the IVS
different near the T/NT interface compared to deep inside the turbulent region? Specically, the interest
is to understand the reason why the radius of the IVS near the jet edge is bigger than inside the rest of the
shear layer? This issue is important in the context of the turbulent entrainment since the nibbling eddy
motions associated with the entrainment are thought to be dictated by the IVS near the T/NT interface.
It has been remarked that only vortices that are approximately described by the Burgers vortex will have
long life times inside a turbulent ow i.e., long lived IVS must have a radius close to the Burgers radius
R R
B
. Therefore the key issue here seems to be the axial stretching rate
0
acting on the axis of the
IVS, given that is of course the same everywhere in the ow.
<>
I
<>
I
0
0.5
y
I
/
25 0 25 50 75
(a)
<(
0
>0)>
I
<
0
>
I
<
0
>
I
0
0.5
1
y
I
/
0 25 50 75
(b)
<(S
ij
S
ij
)
0
>
I
4
5
6
y
I
/
0 25 50 75
(c)
Figure 7.9: Conditional proles of: (a) stretching rate
I
, and mean stretching rate
I
; (b) stretching rate
0
I
,
positive stretching rate (
0
> 0)
I
and mean stretching rate
0
I
, at the axis of the IVS; (c) rate-of-strain norm at
the axis of the IVS

(S
ij
S
ij
)
0
I
I
/ = 20 is one Taylor micro-scale from the T/NT
interface.
The investigation of this issue starts by analysing the conditional proles of stretching rate
I
=
j
S
ij
/(
k
k
)
I
, and mean stretching rate
I
=
i
I
S
ij
I
/ (
k
k
)
I
for the whole ow eld,
in gure 7.9 (a). It is noted from the outset that the total
I
and mean
I
stretching rates are very
different with
I

I
which shows that the mean ow has a negligible inuence on the stretching
rate (for the whole ow eld). Moreover,
I
< 0 is negative in the irrotational region y
I
/ < 0 implying
a prevalence of compression over stretching of the uid elements as the T/NT interface is approached
in agreement with da Silva and Pereira (2008), and increases sharply after the T/NT interface has been
crossed displaying a peak at y
I
/ 5. After this 0.4 for y
I
/ > 30. Since the vorticity norm [[
is roughly constant inside the shear layer, the stretching rate must increase with y
I
for y
I
> 0 like the
rate-of-strain, as described in da Silva and Pereira (2008), provided the alignment between enstrophy
and strain is unchanged. A relatively slow increase in the magnitude of the rate-of-strain norm near
the jet edge was indeed observed independently by da Silva and Pereira (2007a, 2008) and Holzner
et al. (2007), which explains that the enstrophy production increases also slowly after the T/NT interface,
again as shown by da Silva and Pereira (2007a, 2008) and Holzner et al. (2007).
To analyse the stretching rate acting on the axis of the IVS gure 7.9 (b) displays conditional proles
of total axial stretching rate
0
I
=
0
.S.
0
/
2
0
_
I
, positive axial stretching rate (
0
> 0)
I
i.e.,
0
I
for
0
> 0, and mean axial stretching rate
0
I
=
0
I
. S
I
.
0
I
/
2
0
_
I
. The conditional prole of
the total axial stretching rate
0
I
accounts for both positive and negative events, associated with vortex
stretching and vortex compression, respectively, while (
0
> 0)
I
accounts only for vortex stretching,
which is more important than compression, and explains the observed differences between the two
curves i.e., (
0
> 0)
I
>
0
I
. The contribution of the mean eld gradients to the total stretching can
be appreciated in
0
I
. Unlike for the whole turbulent ow,
0
has a non negligible inuence of the
mean eld, as can be attested by the fact that
0
I
and
0
I
have comparable magnitudes. An even
more striking feature of
0
I
as well as of
0
I
, is that it attains its minimum right at y
I
/ 20 i.e.,
exactly at the point where the radius of the IVS is maximum, implying that its seems that the evolution of
the conditional radius of the IVS is indeed caused by the behaviour of
0
near the jet edge. Deep inside
the turbulent region y
I
/ > 40
0
I
,
0
I
, and (
0
> 0)
I
are roughly constant, in contrast with the
region between 0 < y
I
/ < 40, where
0
I
is seen to decay to its minimum.
It remains to understand what causes the decay of
0
near the T/NT interface, to which this investiga-
tion now turns.
0
depends on the rate of strain (S
ij
)
0
and on the vorticity
0
on the axis of the IVS, but
as seen before
0
is roughly constant across the whole shear layer (see gure 7.6 b). As for the strain
acting on the axis of the IVS the gure 7.9 (c) displaying conditional rate-of-strain (S
ij
S
ij
)
0
on the axis
of the IVS shows that this quantity alone cannot explain the observed behaviour of
0
at 0 < y
I
/ < 40,
since (S
ij
S
ij
)
0
at y
I
/ = 20 is similar to (S
ij
S
ij
)
0
for y
I
/ > 40. The observed decay of
0
I
near
the T/NT interface, however, starts by y
I
/ 40. Recall that da Silva and Pereira (2008) and Holzner
et al. (2007) showed that the rate-of-strain for the whole ow eld is also roughly constant inside the
turbulent region. Therefore the decay of
0
around y
I
/ 40 has to be explained by a decrease in the
alignment between
0
and (S
ij
)
0
near the jet edge. A decrease in the alignment between
i
and S
ij
(for
the entire eld) near the T/NT interface was reported in Holzner et al. (2007) and is consistent with the
T/NT interface affecting more the orientation than the magnitude of the vorticity eld described before.
There is however another explanation for the observed decay of
0
near the T/NT interface. Indeed
it has been shown that in isotropic turbulence the axial stretching rate acting on the IVS originates in the
background vorticity structures, and these structures on the other hand lie at the borders of the weak
vorticity or velocity eddy structures, which are themselves relatively free from vorticity in their centres
Jimnez and Wray (1998). Drawing a parallel with the present jet ow the biggest LVS from the jet with
radius of the order of the Taylor micro-scale consists on weak vorticity structures whereas the smaller
LVS are the background vorticity structures. The IVS are then stretched in the borders of the biggest
LVS in the ow. To better visualise the eddies near the T/NT interface Fig. 7.10 shows a sketch of the
ow depicting the LVS and IVS near the jet edge. The IVS deep inside the shear layer are stretched by
the surrounding large scale structures, and the resulting stretching rate imposes their radius R 4.6.
For the IVS near the T/NT interface however, only a fraction of the LVS exist and therefore the overall
stretching rate on the IVS will be smaller than inside the jet. Consequently, it is expected that the
stretching rate acting on the IVS near the jet edge are smaller by a factor
than the stretching rate

inside the jet
0
(y
I
/ = ) =
0
(y
I
/ > ), and this explains the bigger sized IVS found near the
Figure 7.10: Sketch of the (small) intense vorticity structures (IVS) and (big) large vorticity structures (LVS) near
the T/NT interface in the jet. The T/NT interface is dened by the LVS near the jet edge and the IVS appear more
to the interior of the jet shear layer. The average radius of the IVS near the T/NT interface is bigger than deep
inside the jet shear layer.
T/NT interface. In the present results this factor is close to
3/4.
The picture that emerges for the IVS in the jet is the following: the IVS inside the jet are approximately
described by the Burgers vortex model R R
B
= 2 (/
0
)
1/2
, although arguably not as accurately as
in forced isotropic turbulence. In any case the most probable dynamic situation for an IVS inside the
jet consists in an approximate balance between the opposing effects of viscous enstrophy diffusion and
enstrophy production by axial stretching. For the IVS near the jet edge however, the absence of one
fraction of the LVS due to the proximity of the irrotational ow region, decreases the level of total axial
stretching rate imposed on the IVS, compared with the level of stretching found inside the jet. Because
of this the radius of the IVS near the jet edge will grow until they reach a new situation of equilibrium.
Arguably, if a IVS is detrained and lost into the irrotational region, it will tend to grow even more due to
the associated decrease of axial stretching as it moves farther away from the jet centreline. One must
bear in mind that however that since T
T
U
it is clear that the IVS are dissipated in a very small
timescale compared to the time associated with large scale transport in the jet. Thus it is highly unlikely
that the IVS from jets will be observed to decay as the jet evolves downstream.
Finally, notice that in the present work no effort was made to separate the results depending on the
local shape of the T/NT interface surface i.e. whether the interface is convex or concave, such as is done
in e.g. Bisset et al. (2002), where it is shown that the conditional statistics in the convex outer part of the
LVS engulng regions is different from the concave convoluted T/NT interface found near the centreline
of the jet, however this aspect certainly deserves to be investigated.
7.3.4 Lifetime and characteristics of the intense vorticity structures inside the jet
It is possible to make some rough estimates of the characteristics of the IVS inside the shear layer
e.g., the vortex core radius, tangential velocity, vorticity and circulation, as the structures travel down-
stream in the jet. Before doing this, however, it is important to realise that the IVS have a very small life
time compared to the LVS, and that for this reason it will be hard to visualise the scaling laws described
7.4 Summary 113
here as the IVS travel downstream in a jet, since they will be quickly dissipated. Rather the relations
derived here have to be understood as estimates for the existing IVS at a particular downstream location
in the jet. This is seen by comparing the lifetime of the IVS with the characteristic timescale associated
with their downstream transport in the jet. The time needed for an IVS to travel a distance equal to
the integral scale L
11
is T
U
= L
11
/ u, where u is the characteristic mean velocity in the jet, which is
of the order of the centreline mean velocity u u
0
, while the integral scale inside the jet is roughly of
the order of the jet half-width L
11

0.5
(0.4 < L
11
/
0.5
< 0.8 Pope (2000)). On the other hand the
lifetime of the IVS is of the order of the time corresponding to one rotation T
R/u
0
. Given that the
size of the vortex core radius is R , and using u
0
u
, where u
0
is the maximum azimuthal velocity
of the IVS, the result is T
/u
. In a planar jet 0.2 < u
/ u
0
< 0.3 for a range of Reynolds num-
bers between 1500 < Re
H
< 81400 (Deo et al. (2008)), therefore, by taking u
/ u
0
0.25, the result is
T
/T
U
(R/u
0
) ( u
0
/L
11
) 4 (/L
11
) 4Re
3/4
0
, where Re
0
is the Reynolds number of the large scale
eddies Re
0
= u
L
11
/. Clearly the lifetime of the IVS is much shorter than the time associated with their
downstream transport.
The characteristics of the IVS in a planar turbulent jet can be obtained using the theoretical results for
the self-similar plane jet and considering that the IVS are approximately described by the same scaling
laws obtained by Jimnez and Wray (1998) in forced isotropic turbulence. The results described in the
previous section support this view e.g., in forced isotropic turbulence the radius of the IVS is of the order
of the Burgers radius R R
B
and, as seen before, the same is also true in a jet, particularly deep inside
the shear layer. Thus, the streamwise evolution of the vortex core size can be estimated assuming R
R
B
(/
0
)
1/2
. Since the stretching rate is of the order of
0

and given that
/ (Jimnez
and Wray (1998)), and by using the relation for the downstream evolution of the Taylor micro-scale
x
3/4
(Antonia et al. (1980)), one concludes that the IVS radius evolves as R x
5/8
. It is interesting
to recall that in a plane jet the Kolmogorov micro-scale also evolves as x
5/8
as shown by Deo et al.
(2008) and Antonia et al. (1980) which shows that R anywhere in the downstream evolution of the
jet. Given that the Taylor based Reynolds number increases in the jet as Re
= u
/ x
1/4
(Antonia
et al. (1980)) it follows that the circulation evolves as 20Re
1/2
x
1/8
, and the axial vorticity as
Re
1/2
x
9/8
. Finally, the maximum azimuthal velocity evolves as u
0
u
x
1/2
. Naturally it
is expected that, as the T/NT interface is approached, the IVS will tend to have bigger vortex radius, and
smaller axial vorticity, circulation and azimuthal velocity than deep inside the shear layer for the reasons
explained before. The present scaling laws for the IVS could be assessed in a spatially evolving planar
jet.
7.4 Summary
A DNS of a turbulent plane jet at Re
= 120 is used to analyse the intense vorticity structures (IVS)

near the turbulent/non-turbulent (T/NT) interface. The investigation used a tracking procedure to identify
the IVS which is similar (with some modications) to the procedure developed in Jimnez et al. (1993);
Jimnez and Wray (1998). The T/NT interface is roughly dened by the radius of the LVS at the edge
of the jet, which is R
lvs
, while the IVS appear more to the interior of the jet shear layer, at about 5
from the location of the T/NT interface.
Well inside the jet shear layer the IVS characteristics are approximately similar to the IVS observed
in forced isotropic turbulence and several other ows: the mean radius is of the order of the Kol-
mogorov micro-scale R / = 4.6, the tangential velocity is of the order of the root-mean-square velocity
u
0
/u
= 0.76, the mean circulation Reynolds number normalised by the Taylor scale Reynolds number
is Re
/Re
1/2
28, and the mean axial vorticity is

_
0
/
_
Re
1/2
__
= 0.38, where
is the uctuating
vorticity norm. The shape of the PDFs of these quantities are also similar to the corresponding PDFs in
forced isotropic turbulence.
The mean size of the vortex core radius is also very close to the Burgers vortex radius R/R
B
and
the shape of the PDFs of R/R
B
are almost equal in the plane jet and in forced isotropic turbulence. This
indicates that the IVS from the jet are approximately described by the Burgers vortex model i.e., for the
IVS inside a jet the enstrophy production by axial stretching is roughly balanced by the radial viscous
diffusion indicating that the radius of the IVS is stable, as is the case in isotropic turbulence (Jimnez
and Wray (1998)).
The number of detected IVS is roughly constant inside the jet but drops sharply in the region 5 <
y
I
/ < 20. Virtually no IVS exist in the region 0 < y
I
/ < 5 which can be explained by the size of the
IVS being approximately equal to R/ 5 and therefore no IVS can t inside this thin layer.
Conditional statistics in relation to the distance from the T/NT interface were carried out in order
to investigate how the IVS characteristics are affected by the presence of the T/NT interface. The
radius, circulation, vorticity, and tangential velocity are approximately constant inside the jet shear layer.
When approaching the T/NT interface from the interior of the jet (y
I
/ 0) the radius, circulation, and
tangential velocity increase near the jet edge, and display a maximum at about one Taylor micro-scale
from the T/NT interface y
I
/ 20, before dropping again sharply in the region 5 < y
I
/ < 20. The axial
vorticity is approximately constant across the shear layer.
The Burgers vortex is also a good model for the IVS near the jet edge, although not as accurate
as inside the jet shear layer. Indeed for the IVS located at 5 < y
I
/ < 20 R/R
B
0.90. This is
related with the level of axial stretching rate acting on the axis of the IVS
0
, which is roughly constant
inside the shear layer but decreases sharply near the T/NT interface attaining a minimum at y
I
/ 20.
This is exactly the location where the radius, tangential velocity and circulation of the IVS attain their
maximum. The decrease in the magnitude of
0
near the jet edge can be explained by a decrease in the
number of LVS neighbouring the IVS since the stretching imposed on the IVS is known to originate in
the background vorticity at the edges of the LVS. Thus the IVS near the jet edge are are not equilibrium
Burgers vortices, and their mean radius will tend to increase in time. The lifetime of these structures
however, is small compared to the convective timescale making the temporal evolution of the radius
hard to observe in a jet.
The results in the chapter were published in da Silva et al. (in press).
CHAPTER 8
Parallel plane jet simulation
This chapter presents the results obtained from the parallel jet simulation code. Two massive plane jet
simulation are presented.
8.1 Simulations parameters
Two simulations were made using the new parallel DNS code to simulate spatial turbulent plane jets.
Tab. (8.1). summarises their characteristics.
N
x
N
y
N
z
L
x
L
y
L
z
Re
H
H/ noise x/H
DNS0 1280x1280x256 20H 20H 3.99H 6000 30 3% 0.016
DNS1 2128x1568x172 63.8H 47H 5.16H 6000 30 3% 0.030
Table 8.1: Parameters of the parallel turbulent plane jet simulations. Re
H
is the Reynolds number based on the jet
slit width, jet momentum thickness (), x is the grid spacing and noise refers to the noise percentage added to
the average velocity prole, as explained in sec. (3.6.1)
8.1.1 Boundary conditions
An hyperbolic tangent velocity prole based is imposed on the jet inlet.
U =
U
1
+ U
2
2
+
U
1
U
2
2
tanh
_
h
2
0
y
sl
h
_
(8.1)
where U
1
is the jet centreline velocity and U
2
is a superimposed co-ow velocity. The mean y and z
components of the velocity are set to zero at the inlet.
0
is the momentum thickness of the initial shear
layer, h the jets inlet slot width and y
s
l the distance from the centre of the shear layer.
116 Parallel plane jet simulation
y
/
H
1
0.5
0
0.5
1
u/U
1
0 0.5 1
Figure 8.1: Tangent velocity prole
A random velocity eld, exhibiting the statistical characteristics of isotropic turbulence, is then super-
imposed to this prole. The velocity components of this random eld are obtained so to conform to an
energy spectrum given by
E(x) k
s
exp
(0.5s(k/k
0
)
2
)
(8.2)
where k = (k
2
y
+ k
2
z
)
1/2
is the wave-number norm for the (y, z) plane. The values of the exponent s and
peak wave-number k
0
determine the distribution and evolution of energy of the prole: an high k
0
value
to give emphasis to small scales and s 4 for a content typical of freely decaying turbulence. The noise
was imposed only to the shear layer segment of the inlet through a convolution function.
8.1.2 Results DNS0
Fig. (8.2) shows a slice of the domain box where the dimensions in the streamwise and normal
directions can be seen. It can also be seen that the ow exits without problem, conrming the correct
implementation of the outow boundary conditions.
Fig. (8.3) highlights the presence of the LVS in the ow using pressure iso-surfaces. The big initial
billows originated from the Kelvin-Helmholtz instability are present. It can also be see the generation of
secondary vortexes that interact and fragment, increasing the complexity of the ow in the streamwise
direction. This can be further seen in g. (8.4), g. (8.5) and g. (8.6).
Fig. (8.8) shows the growth of the jet half-width (
0.5
) and evolution of the velocity in the centreline.
These were tted to the expressions
0.5
h
= K
u1
_
x
h
+ K
u2
_
(2.58)
_
u
c
(x) u(x = 0, y = )
u
c
u(x = 0, y = )
_
2
= C
u1
_
x
h
+ C
u2
_
(2.59)
repeated here from section 2.2.2.
Fig. (8.8) shows the evolution of the jet half-width (
0.5
) and the centreline velocity. The self-similarity
region can be seen. It is also patent that the jet transition to turbulence is relatively late. The slope
values can be seen in tab. (8.2) and although near those expected are still smaller than those reported
in the literature.
8.1 Simulations parameters 117
Figure 8.2: DNS0: vertical slice coloured by velocity norm and displaying velocity streamlines. It can be seen the
ow exits properly at the outlet.
Figure 8.3: DNS0: Slanted side-view of iso-contour surfaces for pr = 0.1 (white).
Figure 8.4: DNS0: Slanted side-view of iso-contour surfaces for || = 30 (orange) and pr = 0.1 (white).
Figure 8.5: DNS0: Side and top-down views of iso-contour surfaces for || = 30 (orange) and pr = 0.1 (white).
Figure 8.6: DNS0: Side and top-down views of iso-contour surfaces for Q = 60 (red), Q = 60 (blue) and
pr = 0.1 (white).
Figure 8.7: DNS0: transition region zoom. a) shows pr = 0.1. In b) the pressure iso-surfaces were made
translucent to better highlight the presence of (orange, || = 30|. In c) the same approach was taken to
showcase the Q criterion: red signals Q = 60 and blue Q = 60.
DNS0
d
0
.
5
/
H
0.5
1
1.5
2
x/H
0 5 10 15 20
DNS0
[
(
U
1
-
U
2
L
)
]
/
[
(
U
c
-
U
2
)
]
-
2
0
1
2
3
x/H
0 5 10 15 20
Figure 8.8: Half-width growth (left) and centreline velocity evolution (right) for DNS0. U
2L
is the local co-ow
velocity. The values for the tting are in tab. (8.2)
8.1.3 Results DNS1
Figure 8.9: DNS1: vertical slice coloured by velocity norm and displaying velocity streamlines. The oscillation
behaviour perceived in the streamlines might indicate that the jet is conned and either a higher co-ow or an
increase in the vertical domain width should be pursued.
In DNS1 the domain box is uneven in the streamwise and normal directions (g. (8.2)). No problem
arise in regard the outow as can be seen by inspecting the velocity streamlines presented.
Table 8.2 summarises the results for the
0.5
and u
c
for the simulations done and for several other
simulations presented in the literature. The location of the jet virtual origin is expected to vary due to
the high dependence on the inlet conditions (see Stanley and Sarkar (2000b)). Small departures on the
variation rates are also expected, although for the present DNS1 the values presented are below than
those found in the literature.
Figure 8.10: DNS1: Side and top-down views of iso-contour surfaces for || = 30 (orange) and pr = 0.1 (white).
Figure 8.11: DNS1: Side and top-down views of iso-contour surfaces for Q = 60 (red), Q = 60 (blue) and
pr = 0.1 (white).
DNS1
d
0
.
5
/
H
0
1
2
3
4
x/H
0 20 40 60
slope = 0.12 x/H = [7,50]
slope = 0.17, x/H=[7,20]
DNS1
[
(
U
1
-
U
2
L
)
]
/
[
(
U
c
-
U
2
)
]
-
2
0
2
4
6
x/H
0 20 40 60
Figure 8.12: Centreline velocity and half-width growth for DNS1. The values for the tting are in tab. (8.2)
source K
u1
K
u2
C
u1
C
u1
DNS0 0.087 -1.28 0.119 1.75
DNS1 0.064 1.92 0.13 3.24
da Silva and Mtais (2002a) 0 0.089 2.10 0.165 -0.85
da Silva and Mtais (2002a) 1 0.108 -2.10 0.20 -2.75
da Silva and Mtais (2002a) 2 0.089 0.80 0.165 -0.25
Stanley et al. (2002) 0.092 2.62 0.201 1.23
Ramparian and Chandrasekhara (1985) 0.110 -1.00 0.093 -1.60
Browne et al. (1982) 0.104 -5.00 0.143 -9.00
Thomas and Prakash (1991) 0.110 0.14 0.220 -1.20
Thomas and Chu (1989) 0.110 0.14 0.220 -1.19
Gutmark and Wygnansky (1976) 0.100 -2.00 0.189 -4.72
Table 8.2: Parameters of the simulations
8.2 Parallel performance
8.2.1 Initial version vs new version
As already described the parallel version of the code took departure from the implementation of
da Silva and Mtais (2002a). Because the nal code introduces several improvements over the original
version, a rst comparison, using sequential runs, is made. Four versions are thus compared: the
original version (V 0.0), a naive parallel implementation (V 1.0), a parallel version with index re-ordering
(V 1.5) and the nal version (V 2.0), with index re-ordering and using the FFTW library for solving the
FFTs. The rst three versions use Temperton (1983b,a) for calculating the FFTs.
V2.0
V1.5
V1.0
S
p
e
e
d
u
p

v
s

V
0
.
0
0
5
10
N
proc
0 2 4 6 8
Figure 8.13: Speedup against V0.0 implementation. Compiled with gfortran 4.4, openmpi 1.4.1, using a 2x Intel
Quadcore Xeon X5460 @ 3.16GHz and having 32GB DDR2 RAM running GNU/Linux 2.6.32-3.
(t(V1.0)/t(V1.5))
NP 320x320x72 640x640x144 640x640x360 960x960x160
1 1.68 2.13 2.24 2.07
2 1.06 2.43 1.82 2.45
4 1.20 1.94 1.97 2.05
8 1.29 2.11 2.08 2.07
Table 8.3: Comparison of gains between version 1.0 and 1.5. Compiled with gfortran 4.4, openmpi 1.4.1, using a
2x Intel Quadcore Xeon X5460 @ 3.16GHz and having 32GB DDR2 RAM running GNU/Linux 2.6.32-3.
The graphic 8.13 shows the speedup measured against the initial sequential version. It can be seen,
for a test grid of 640 640 144 a constant performance increase between the different implementation
strategies.
Comparing table 8.3 and 8.4, its manifest that the highest increase in performance was achieved by
the index re-ordering strategy (gains in table 8.3). The original version kept an ijk ordering through out
(t(V1.5)/t(V2.0))
NP 320x320x72 640x640x144 640x640x360 960x960x160
1 1.39 1.69 1.56 1.26
2 1.45 1.42 1.52 1.07
4 1.43 1.54 1.61 1.10
8 1.40 1.51 1.57 1.11
Table 8.4: Comparison of gains between version 1.5 and 2.0. Compiled with gfortran 4.4, openmpi 1.4.1, using a
2x Intel Quadcore Xeon X5460 @ 3.16GHz and having 32GB DDR2 RAM running GNU/Linux 2.6.32-3.
the code, resorting to a local jk re-ordering operation per i line when calculating the FFT. The change
from the Temperton FFT algorithm to the use of the FFTW library was the main responsible for the
performance increase from V1.5 to V2.0.
320x320x72
640x640x144
640x640x360
960x960x144
S
p
e
e
d
u
p

V
2
.
0
0
1
2
3
4
N
proc
0 2 4 6 8
Figure 8.14: Speedup for V2.0. Comparison of V2.0 behaviour versus problem size increase.
8.2.2 Interconnect inuence
A small test was done to access the impact of the interconnect type on parallel performance. This
test was limited by resource availability.
Results are shown in table 8.5. Runs where made in the Iniciativa Nacional Grid (INGRID) cluster
which has both types of interconnect: InniBand SDR and GigE.
I/O was absent from the test runs to eliminate any inuence on the runs (so the ratios presented are
only related to the calculation time). As expected, the interconnect type plays a crucial role in the code
performance.
NPROCS Nodes t GigE / t InniBand
1280 1024 202 16 2 1.99
32 4 2.31
64 8 2.42
1280 1024 404 32 4 2.49
64 8 2.58
Table 8.5: Interconnect performance dependence. GigE vs. InniBand. InniBand connect of the SDR type.
8.2.3 Master I/O vs MPI-IO
To compare the performance of both I/O strategies exposed in sec. 4.7 some tests were done using
either a single machine or runs in the INGRID cluster. They are presented in table 8.6. The results show
consistency except for the combination of 32/4 in the table. This should be explained by the presence
other processes activity on the system on the machine during the test which contaminated the result.
size NPROCS / Nodes Mast. vs. MPI-IO no-I/O vs. Mast. no-I/O vs. MPI-IO
single 960 960 160 8 / 1 0.93
1264 1008 202 8 / 1 1.05
cluster 1280 1064 202 16 / 2 1.23 0.5 0.62
1280 1064 404 32 / 4 17.87 0.05 0.95
1280 1064 404 64 / 8 1.67 0.58 0.98
Table 8.6: IO performance: t of Master writes vs. t of MPI-IO. Test done with 10 output cycles of velocity eld to
disk. Last two columns show performance loss against running without I/O.
8.3 Summary
Two spatial turbulence plane jet simulations were made and their results reported. The performance
analysis shows the advantages brought by the several different strategies, with emphasis on the use of
the FFT library FFTW, index re-ordering and MPI-IO. It also highlights the impact of the presence high
performance interconnect in this type of codes.
CHAPTER 9
Conclusion
A summary of conclusions is presented and some considerations about future work directions are
presented.
9.1 Main Contributions
A tracking algorithm was implemented to educe intense vorticity structures (IVS) from turbulent ows.
This algorithm was validated against an homogeneous isotropic turbulence (HIT) database and then
used to study the IVS in a turbulent plane jet, using a temporal DNS simulation with Re
= 120. The role

of these structures in the presence of the T/NT interface was also studied with the same database.
The coherent vortices near the turbulent/non-turbulent interface can be divided into large vortical
structures (LVS) maintained by the mean shear, and intense vorticity structures (IVS) created by the
background turbulence.
The characteristic vorticity jump of the T/NT interface as well as its thickness is dictated by the radial
vorticity distribution of the bigger LVS near the T/NT interface. Specically the radius of these LVS is
equal to the Taylor micro-scale and explains why this is also the average thickness of the T/NT interface.
The coherent vortices in the proximity of the T/NT interface are preferentially aligned with the tangent to
the T/NT interface and are responsible for the viscous dissipation of kinetic energy observed outside the
turbulent region.
It is also argued that the nibbling eddy motions responsible for the turbulent entrainment are the
manifestation of the enstrophy viscous diffusion that exists near the edges of the LVS sitting near the
T/NT interface.
The T/NT interface is roughly dened by the radius of the LVS at the edge of the jet, which is R
lvs
,
while the IVS appear more to the interior of the jet shear layer, at about 5 from the location of the T/NT
128 Conclusion
interface.
Well inside the jet shear layer the IVS characteristics are approximately similar to the IVS observed
in forced isotropic turbulence and several other ows: the mean radius is of the order of the Kol-
mogorov micro-scale R / = 4.6, the tangential velocity is of the order of the root-mean-square velocity
u
0
/u
= 0.76, the mean circulation Reynolds number normalized by the Taylor scale Reynolds number
is Re
/Re
1/2
28, and the mean axial vorticity is

_
0
/
_
Re
1/2
__
= 0.38, where
is the uctuating
vorticity norm. The shape of the PDFs of these quantities are also similar to the corresponding PDFs in
forced isotropic turbulence.
The mean size of the vortex core radius is also very close to the Burgers vortex radius R/R
B
and
the shape of the PDFs of R/R
B
are almost equal in the plane jet and in forced isotropic turbulence. This
indicates that the IVS from the jet are approximately described by the Burgers vortex model i.e., for the
IVS inside a jet the enstrophy production by axial stretching is roughly balanced by the radial viscous
diffusion indicating that the radius of the IVS is stable, as is the case in isotropic turbulence (Jimnez
and Wray (1998)).
The number of detected IVS is roughly constant inside the jet but drops sharply in the region 5 <
y
I
/ < 20. Virtually no IVS exist in the region 0 < y
I
/ < 5 which can be explained by the size of the
IVS being approximately equal to R/ 5 and therefore no IVS can t inside this thin layer.
Conditional statistics in relation to the distance from the T/NT interface were carried out in order
to investigate how the IVS characteristics are affected by the presence of the T/NT interface. The
radius, circulation, vorticity, and tangential velocity are approximately constant inside the jet shear layer.
When approaching the T/NT interface from the interior of the jet (y
I
/ 0) the radius, circulation, and
tangential velocity increase near the jet edge, and display a maximum at about one Taylor micro-scale
from the T/NT interface y
I
/ 20, before dropping again sharply in the region 5 < y
I
/ < 20. The axial
vorticity is approximately constant across the shear layer.
The Burgers vortex is also a good model for the IVS near the jet edge, although not as accurate as
inside the jet shear layer. Indeed for the IVS located at 5 < y
I
/ < 20 R/R
B
0.90. This is related
with the level of axial stretching rate acting on the axis of the IVS
0
, which is roughly constant inside
the shear layer but decreases sharply near the T/NT interface attaining a minimum at y
I
/ 20. This is
exactly the location where the radius, tangential velocity and circulation of the IVS attain their maximum.
The decrease in the magnitude of
0
near the jet edge can be explained by a decrease in the number of
LVS neighboring the IVS since the stretching imposed on the IVS is known to originate in the background
vorticity at the edges of the LVS. Thus the IVS near the jet edge are are not equilibrium Burgers vortices,
and their mean radius will tend to increase in time. The lifetime of these structures however, is small
compared to the convective timescale making the temporal evolution of the radius hard to observe in a
jet.
A new parallel spatial turbulent plane jet code was implemented using a modular strategy which al-
lows for easiness in future extension. This code increases the research capabilities of the scientic com-
munity, being already in use with some of the researchers at Laboratrio de Energia e Fluidos (LASEF).
In this respect there are already versions capable of passive scalar, dynamic boundary conditions (Lopes
9.1 Main Contributions 129
et al. (2010)). Compressibility capability is being added at the moment of writing.
Two big turbulent plane jet simulations were made. They will allow the advancing of the studies about
the T/NT interface, applying the methodology used to a spatial evolving turbulent plane jet ow.
Two papers were published in international peer reviewed journals, da Silva and dos Reis (2011) and
da Silva et al. (in press).
Bibliography
G. Amdahl. Validity of the single processor approach to achieving large scale computing capabilities. In
AFIPS Conference Proceedings, volume 30, pages 483485. 1967. (One citation in page 66.)
R. K. Anand, B. J. Boersma, and A. Agrawal. Detection of turbulent/non-turbulent interface for an ax-
isymmetric turbulent jet: evaluation of known criteria and proposal of a new criterion. Expts. Fluids,
997-1007:479(6), 2009. (Two citations in pages 75 and 100.)
R. A. Antonia, B. R. Satyaprakash, and A. K. M. F. Hussain. Measurements of dissipation rate and some
other characteristics of turbulent plane and circular jets. Phys. Fluids, 23(4):695700, 1980. (One
citation in page 113.)
W. Ashust, A. Kerstein, R. Kerr, and C. Gibson. Alignment of vorticity and scalar gradient with strain rate
in simulated navier-stokes turbulence. Phys. Fluids, 30:2343, 1987. (One citation in page 33.)
D. K. Bisset, J. C. R. Hunt, and M. M. Rogers. The turbulent/non-turbulent interface bounding a far wake.
J. Fluid Mech., 451:383410, 2002. (14 citations in pages 29, 31, 32, 33, 34, 75, 79, 80, 81, 82, 100,
103, 104, and 112.)
L. Bradbury. The structure of a self-preserving turbulent plane jet. J. Fluid Mech., 23:3164, 1965. (One
P. Bradshaw. Turbulence: the chief outstanding difculty of our subject. Expts. Fluids, 16:203216,
1994. (One citation in page 7.)
G. Brown and A. Roshko. On Density Effects and Large Structure in Turbulent Mixing Layers. J. Fluid
Mech., 64:775816, 1974. (One citation in page 29.)
L. J. S. Browne, R. A. Antonia, S. Rajagopalan, and A. J. Chambers. Interaction region of a two-
dimensional turbulent plane jet in still air. In R. Dumas and L. Fulachier, editors, Structure of complex
turbulent shear ow, IUTAM Symp. Marseille. Springer, 1982. (One citation in page 123.)
132 BIBLIOGRAPHY
R. Camussi and G. Guj. Experimental analysis of intermittent coherent vortices in the near eld of a
high re turbulent jet. Phys. Fluids, 11(2):423431, 1999. (One citation in page 27.)
C. Canuto, M. Hussaini, A. Quarteroni, and T. Zang. Spectral Methods in Fluid Dynamics. Springer-
Verlag, 1987. (Two citations in pages 42 and 49.)
D. J. Carruthers and J. C. R. Hunt. Velocity uctuations near an interface between a turbulent region and
a stably stratied layer. J. Fluid Mech., 165:475501, 1986. (One citation in page 29.)
A. Celani. The frontiers of computing in turbulence: challenges and perspectives. Journal of Turbulence,
8(34), 2007. (Two citations in pages 38 and 39.)
P. Chakraborty, S. Balachandar, and R. Adrian. On the relationships between local vortex identication
schemes. J. Fluid Mech., 535:189214, 2005. (One citation in page 19.)
M. Chong, E. Perry, and B. Cantwell. A general classication of three-dimensional ow elds simulations
of turbulence. Phys. Fluids, 2(5):765777, 1990. (One citation in page 19.)
G. Corcos and F. Sherman. The mixing layer: deterministic models of a turbulent ow. part 1 introduction
and the two-dimensional ow. J. Fluid Mech., 139:2965, 1984. (One citation in page 30.)
S. Corrsin and A. L. Kistler. Free-stream boundaries of turbulent ows. Technical Report TN-1244,
NACA, 1955. (Seven citations in pages 29, 31, 32, 79, 80, 87, and 92.)
S. C. Crow and F. H. Champagne. Orderly structure in jet turbulence. J. Fluid Mech., 48:547591, 1971.
(One citation in page 27.)
C. da Silva. The role of coherent structures in the control and interscale interactions of round, plane and
coaxial jets. PhD thesis, Institut National Politechnique de Grenoble, 2001. (One citation in page 27.)
C. da Silva, R. J. Reis, and J. C. Pereira. The intense vorticity structures near the turbulent/nonturbulent
interface in a jet. J. Fluid Mech., in press. (Three citations in pages 3, 114, and 129.)
C. B. da Silva. The behavior of subgrid-scale models near the turbulent/nonturbulent interface in jets.
Phys. Fluids, 21:081702, 2009. (Seven citations in pages 32, 75, 80, 81, 100, 103, and 104.)
C. B. da Silva and R. N. dos Reis. The role of coherent vortices near the turbulent/nonturbulent interface
in a planar jet. Phil. Trans. Roy. Soc. A, 369:738753, 2011. (Eight citations in pages 3, 31, 34, 79,
93, 105, 106, and 129.)
C. B. da Silva and O. Mtais. On the inuence of coherent structures upon interscale interactions in
turbulent plane jets. J. Fluid Mech., 473:103145, 2002a. (Six citations in pages 24, 28, 39, 84, 123,
and 124.)
C. B. da Silva and O. Mtais. Vortex control of bifurcating jets: a numerical study. Phys. Fluids, 14(11):
37983819, 2002b. (One citation in page 39.)
BIBLIOGRAPHY 133
C. B. da Silva and J. C. F. Pereira. Enstrophy, strain and scalar gradient dynamics across the turbulent-
nonturbulent interface in jets. In Advances in Turbulence XI - 11th Euromech ETC, Porto, 2007a. (Four
citations in pages 80, 91, 92, and 110.)
C. B. da Silva and J. C. F. Pereira. Analysis of the gradient-diffusion hypothesis in large-eddy simulations
based on transport equations. Phys. Fluids, 19:035106, 2007b. (Two citations in pages 32 and 95.)
C. B. da Silva and J. C. F. Pereira. Invariants of the velocity-gradient, rate-of-strain, and rate-of-rotation
tensors across the turbulent/nonturbulent interface in jets. Phys. Fluids, 20:055101, 2008. (16 citations
in pages 29, 32, 75, 80, 81, 82, 87, 89, 90, 91, 98, 100, 104, 105, 110, and 111.)
C. B. da Silva and J. C. F. Pereira. Erratum: "invariants of the velocity-gradient, rate-of-strain, and rate-
of-rotation tensors across the turbulent/nonturbulent interface in jets" [phys. uids, 20, 055101, 2008].
Phys. Fluids, 21:019902, 2009. (Four citations in pages 32, 80, 87, and 105.)
C. B. da Silva and R. R. Taveira. The thickness of the turbulent/nonturbulent interface is equal to the
radius of the large vorticity structures near the edge of the shear layer. Phys. Fluids, 22:121702, 2010.
(Seven citations in pages 29, 32, 34, 81, 103, 104, and 105.)
C. B. da Silva, R. dos Reis, and J. C. F. Pereira. On the differences between production and dissipation
mechanisms of enstrophy and scalar gradient across the turbulent/non-turbulent interface in jets. In
6th Turbulent Shear Flow Phenomena, Munich, 2007. (Two citations in pages 89 and 92.)
Y. Dai, T. Kobayashi, and N. Taniguchi. Large eddy simulation of plane turbulent jet ow using a new
outow velocity boundary condition. JSME Intl J. B Fluids and Thermal Engng, 37:242253, 1994.
I. Danaila and J. Boersma. Direct numerical simulation of bifurcating jets. Phys. Fluids, 12(5):1255
1257, 2000. (One citation in page 27.)
P. A. Davidson. Turbulence, an introduction for scientists and engineers. Oxford University Press, 2004.
(Five citations in pages 7, 15, 23, 24, and 100.)
R. Deo, J. Mi, and G. Nathan. The inuence of nozzle-exit geometric prole on statistical properties of
a turbulent jet. Experimental Thermal and Fluid Science, 32(2):545559, 2007. (One citation in page
28.)
R. C. Deo, J. Mi, and G. J. Nathan. The inuence of reynolds number on a plane jet. Phys. Fluids, 20:
A. Dewan, M. Pathak, and K. A. Dass. A survey of selected literature on important ow properties and
computational uid dynamics treatments of incompressible turbulent plane and round jets in quiescent
ambient. Indian Journal of Engineering & Materials Sciences, 13:180194, 2005. (One citation in page
39.)
W. Dham and P. Dimotakis. Measurements of entrainment and mixing in turbulent jets. AIAAJ, 25:
134 BIBLIOGRAPHY
P. E. Dimotakis. The mixing transition in turbulent ows. J. Fluid Mech., 409:69.98, 2000. URL http:
dx.doi.org/10.1017/S0022112099007946. (Two citations in pages 21 and 31.)
S. Douady, Y. Couder, and M. E. Brachet. Direct observation of the intermittency of intense vortex
laments in turbulence. Phys. Review Lett., 67:983, 1991. (One citation in page 16.)
J. A. Ferr, J. C. Mumford, A. M. Savill, and F. Giralt. Three-dimensional large-eddy motions and ne-
scale activity in a plane turbulent wake. J. Fluid Mech., 210:371414, 1990. (One citation in page
31.)
J. Ferziger and M. Peri ` c. Computational methods for uid dynamics. Springer-Verlag, 1996. (Two
citations in pages 42 and 49.)
M. Frigo and S. G. Johnson. The design and implementation of FFTW3. Proceedings of the IEEE, 93
(2):216231, 2005. Special issue on Program Generation, Optimization, and Platform Adaptation.
B. Ganapathisubramani, K. Lakshminarasimhan, and N. T. Clemens. Investigation of three-dimensional
structure of ne scales in a turbulent jet by using cinematographic stereoscopic particle image ve-
locimetry. J. Fluid Mech., 598:141175, 2008. (Two citations in pages 33 and 84.)
R. P. Garg, J. H. Ferziger, and S. G. Monismith. Hybrid spectral nite difference simulations of stratied
turbulent ows on distributed memory architectues. Int. J. Numer. Methods Fluids, 24:11291158,
W. K. George and H. J. Hussein. Locally axisymmetric turbulence. J. Fluid Mech., 233:123, 1991. (One
J. Gustafson. Reevaluating amdahls law. Communications of the ACM, 31:532533, 1988. (One citation
in page 67.)
E. Gutmark and C.-M. Ho. Preferred modes and the spreading rates of jets. Phys. Fluids, 26(10):
E. Gutmark and I. Wygnansky. The planar turbulent jet. J. Fluid Mech., 73:465495, 1976. (One citation
in page 123.)
C. Hamman, R. M. Kirby, and M. Berzins. Parallelization and scalability of a spectral element channel
ow solver for incompressible navier-stoke equations. Concurrency and computation: pratice and
experience, 19:14031422, 2007. (One citation in page 40.)
C.-M. Ho and P. Huerre. Perturbed free shear layers. Annu. Rev. Fluid Mech., 16:356425, 1984. (Three
citations in pages 25, 27, and 28.)
M. Holzner, A. Liberzon, N. Nikitin, W. Kinzelbach, and A. Tsinober. Small-scale aspects of ows in
proximity of the turbulent/nonturbulent interface. Phys. Fluids, 19:071702, 2007. (Eight citations in
pages 32, 80, 87, 88, 90, 92, 110, and 111.)
BIBLIOGRAPHY 135
M. Holzner, A. Liberzon, B. Luthi, N. Nikitin, W. Kinzelbach, and A. Tsinober. A lagrangian investigation
of the small-scale features of turbulent entrainment through particle tracking and direct numerical
simulation. J. Fluid Mech., 598:465475, 2008. (Seven citations in pages 32, 80, 87, 90, 91, 92,
and 93.)
M. Holzner, B. Luthi, A. Liberzon, A. Tsinober, and W. Kinzelbach. The local entrainment velocity is a
viscous quantity. In Turbulence, Heat and Mass Transfer 6 (THMT), Rome, 2009. (Three citations in
pages 32, 80, and 89.)
K. Horiuti. A classication method for vortex sheet and tube structures in turbulent ows. Phys. Fluids,
13:3756, 2001. (Two citations in pages 19 and 84.)
K. Horiuti and T. Ozawa. Multimode strecthed spiral vortex and nonequilibrium energy spectrum in
homogeneous shear ow turbulence. Phys. Fluids, 23, 2010. (One citation in page 20.)
K. Horiuti and Y. Takagi. Identication method for vortex sheet structures in turbulent ows. Phys. Fluids,
17:121713, 2005. (Two citations in pages 20 and 84.)
S. Hoyas and J. Jimenez. Scaling of the velocity uctuations in turbulent channels up to re
= 2003.
PoF, 18(011702), 2006. (Three citations in pages 40, 65, and 71.)
J. C. R. Hunt, A. A. Wray, and P. Moin. Eddies, stream, and convergence zones in turbulent ows.
Annual research briefs, Center for Turbulence Research, Stanford, 1988. (One citation in page 19.)
J. C. R. Hunt, I. Eames, and J. Westerweel. Mechanics of inhomogeneous turbulence and interfacial
layers. J. Fluid Mech., 554:499519, 2006. (Three citations in pages 79, 81, and 88.)
J. C. R. Hunt, I. Eames, and J. Westerweel. Vortical interactions with interfacial shear layers. In
Y. Kaneda, editor, Proceedings of IUTAM conference on Computational Physics and new perspec-
tives in turbulence, Nagoya, Sept 2006. Springer Science, Berlin, 2008. (Four citations in pages 29,
34, 79, and 81.)
J. C. R. Hunt, I. Eames, and J. Westerweel. Eddy dynamics near sharp interfaces and in straining ows.
In Proc. Euro. Conf. Maths. Ind., July 2008. Springer, 2009. (Two citations in pages 79 and 81.)
J. C. R. Hunt, I. Eames, J. Westerweel, P. A. Davidson, S. Voropayev, J. Fernando, and M. Braza. Thin
shear layers - the key to turbulence structure. Journal of Hydro-environment Research, 4:7582,
2010. (Three citations in pages 34, 81, and 88.)
F. Hussain. Coherent structures and turbulence. J. Fluid Mech., 173:303356, 1986. (One citation in
page 28.)
J. Jeong and F. Hussain. On the Identication of a Vortex. J. Fluid Mech., 285:6994, 1995. (Two
J. Jimenez. Kinematic alignment effects in turbulent ows. Phys. Fluids, 4:652654, 1992. (One citation
in page 18.)
136 BIBLIOGRAPHY
J. Jimnez. Computing high-reynolds-number turbulence: will simulations ever replace experiments?
Journal of Turbulence, 4(22), 2003. doi: 10.1088/1468-5248/4/1/022. URL http://dx.doi.org/10.
1088/1468-5248/4/1/022. (Two citations in pages 37 and 39.)
J. Jimnez. Turbulence and vortex dynamics, 2004. (One citation in page 10.)
J. Jimnez and A. Wray. On the characteristics of vortex laments in isotropic turbulence. J. Fluid Mech.,
373:255285, 1998. (16 citations in pages 33, 34, 72, 73, 84, 85, 96, 97, 98, 99, 100, 107, 111, 113,
114, and 128.)
J. Jimnez, A. Wray, P. Saffman, and R. Rogallo. The structure of intense vorticity in isotropic turbulence.
J. Fluid Mech., 255:6590, 1993. (10 citations in pages 33, 72, 73, 84, 85, 96, 98, 99, 103, and 113.)
Y. Kaneda and T. Ishihara. High-resolution direct numerical simulation of turbulence. Journal of Turbu-
lence, 7(20):117, 2006. (One citation in page 39.)
S.-J. Kang, M. Tanahashi, and T. Miyauchi. Dynamics of ne scale eddy clusters in turbulent channel
ows. Journal of Turbulence, 8(52):119, 2008. (Three citations in pages 33, 84, and 98.)
D. A. Kennedy and S. Corrsin. Spectral atness factor and intermittency in turbulence and in non-linear
noise. J. Fluid Mech., 10:366370, 1961. (One citation in page 7.)
S. Kida and H. Miura. Identication and Analysis of Vortical Structures. Eur. J. Mech. B/Fluids, 17(4):
J. Kim and P. Moin. Application of a fractional-step method to incompressible navier-stokes equations.
J. Comp. Phys., 59:308323, 1985. (One citation in page 48.)
A. N. Kolmogorov. Local structure of turbulence in an incompressible uid for very large Reynolds
numbers. Dokl. Aka. Nauk SSSR, 30:301305, 1941. (One citation in page 12.)
H. Le and P. Moin. An improovement of fractional-step methods for the incompressible navier-stokes
equations. J. Comp. Phys., 92:369379, 1991. (One citation in page 48.)
S. K. Lele. Compact nite difference schemes with spectral-like resolution. J. Comp. Phys., 103:1542,
M. Lesieur. Turbulence in uids, third ed. Kluwer, 1997. (Three citations in pages 7, 15, and 16.)
Y. Li, E. Perlman, M. Wan, Y. Yang, C. Meneveau, R. Burns, S. Chen, A. Szalay, and G. Eyink. A public
turbulence database cluster and applications to study lagrangian evolution of velocity increments in
turbulence. Journal of Turbulence, page N31, 2008. doi: 10.1080/14685240802376389. URL http:
//www.tandfonline.com/doi/abs/10.1080/14685240802376389. (One citation in page 71.)
D. C. Lopes, R. J. Reis, C. B. da Silva, and J. C. Pereira. Effect of initial conditions in the far eld of
spatially developing turbulent planar jets. In ECCOMAS-CFD 2010. ECCOMAS, 2010. (One citation
in page 128.)
BIBLIOGRAPHY 137
P. Luchini and M. Quadrio. A low-cost parallel implementation of direct numerical simulation of wall
turbulence. JCP, 211:551571, 2006. (Two citations in pages 40 and 60.)
J. Mathew and A. Basu. Some characteristics of entrainment at a cylindrical turbulent boundary. Phys.
Fluids, 14(7):20652072, 2002. (Five citations in pages 29, 31, 80, 100, and 103.)
J. Mellado, L. Wang, and N. Peters. Gradient trajectory analysis of a scalar eld with external intermit-
tency. J. Fluid Mech., 626:333365, 2009. (One citation in page 83.)
MPI-2: Extensions to the Message-Passing Interface. Message Passing Interface Forum, 1997. URL
http://www-unix.mcs.anl.gov/mpi/mpi-standard/mpi-report-2.0/mpi2-report.htm. (One cita-
tion in page 57.)
A. Michalke. On spatially growing disturbances in an inviscid shar layer. J. Fluid Mech., 23:521544,
P. Moin and J. Kim. Numerical investigations of turbulent channel oweld. J. Fluid Mech., 118, 1982.
P. Moin and K. Mahesh. Direct numerical simulation: a tool in turbulence research. Annu. Rev. Fluid
Mech., 30:539578, 1998. (Three citations in pages 38, 39, and 40.)
H. Mouri, A. Houri, and Y. Kawashima. Laboratory experiments for intense vortical structures in turbu-
lence velocity elds. Phys. Fluids, 19:055101, 2007. (Two citations in pages 84 and 98.)
A. Obukhov. Spectral energy distribution in turbulent ow. Dokl. Akad. Nauk. USSR, 1:2224, 1941.
M. Olsson and L. Fuchs. A simple boundary condition for unbounded hyperbolic ows. Phys. Fluids,
8(8):21252137, 1996. (One citation in page 27.)
I. Orlanski. A simple boundary condition for unbounded hyperbolic ows. J. Comp. Phys., 21:251269,
S. Orszag and G. Patterson. Numerical simulation of three-dimensional homogeneous isotropic turbu-
lence. Phys. Rev. Lett., 28:7679, 1972. (Two citations in pages 37 and 39.)
R. Pelz. The parallel fourie pseudospectral method. JCP, 92:269312, 1992. (One citation in page 61.)
O. M. Phillips. The irrotational motion outside a free turbulent boundary. Proc. Camb. Phil. Soc., 51:220,
1955. (Two citations in pages 29 and 31.)
S. B. Pope. Turbulent Flows. Cambridge University Press, 2000. (Five citations in pages 14, 24, 25, 38,
and 113.)
A. Povitsky. Parallelization of the pipelined thomas algorithm. Technical report, NASA, 1998. (One
138 BIBLIOGRAPHY
R. Ramparian and M. S. Chandrasekhara. LDA measurements in plane turbulent jets. ASME Fluids
Eng., 107:264271, 1985. (One citation in page 123.)
O. Reynolds. An experimental investigation of the circunstances which determine whether the motion
of water shal be direct or sinuous, and of the law of resistance in parallel channels. Phil. Trans. Roy.
Soc., 174:935982, 1883. (One citation in page 8.)
W. C. Reynolds. Large-scale instabilities of turbulent wakes. J. Fluid Mech., 54:481488, 1972. (Two
C. L. Ribault, S. Sarkar, and S. Stanley. Large eddy simulation of a plane jet. Phys. Fluids, 11(10):
L. Richardson. Weather prediction by numerical process. Cambridge University Press, 1922. (One
S. Saddoughi and S. Veeravalli. Local isotropy in turbulent boundary layers at high Reynolds number. J.
Fluid Mech., 268:333372, 1994. (One citation in page 14.)
P. Sagaut. Introduction la simulation des grandes chelles pour les coulements de uide incompress-
ible. Springer-Verlag, 1998. (Two citations in pages 15 and 17.)
H. Sato. The stability and transition of a two-dimensional jet. J. Fluid Mech., 7:5380, 1960. (One
E. D. Siggia. Numerical Study of Small-Scale Intermittency in Three-Dimensional Turbulence. J. Fluid
Mech., 107:375406, 1981. (Three citations in pages 16, 33, and 84.)
S. Stanley and S. Sarkar. Inuence of nozzle conditions and discrete forcing on turbulent planar jets.
AIAA J., 38:16151623, 2000a. (One citation in page 50.)
S. Stanley and S. Sarkar. Inuence of nozzle conditions and discrete forcing on turbulent planar jets.
AIAA J., 38:16151623, 2000b. (One citation in page 121.)
S. Stanley, S. Sarkar, and J. P. Mellado. A study of the oweld evolution and mixing in a planar turbulent
jet using direct numerical simulation. J. Fluid Mech., 450:377407, 2002. (Six citations in pages 24,
39, 50, 82, 84, and 123.)
M. Tanahashi, S. Iwase, and T. Miyauchi. Appearance and alignment with strain rate of coherent ne
scale eddies in turbulent mixing layer. Journal of Turbulence, 2(6):117, 2001. (Four citations in pages
33, 84, 98, and 99.)
M. Tanaka and S. Kida. Characterization of vortex tubes and sheets. Phys. Fluids, 5:20792082, 1993.
C. Temperton. Fast mixed-radix real Fourier transforms. J. Comp. Phys., 52:340350, 1983a. (One
BIBLIOGRAPHY 139
C. Temperton. Self-sorting mixed-radix fast Fourier transforms. J. Comp. Phys., 52:123, 1983b. (One
H. Tennekes and J. L. Lumley. A rst course in turbulence. The MIT Press, 1972. (One citation in page
90.)
F. Thomas and V. Goldschmidt. Structural characteristics of a developing turbulent planar jet. J. Fluid
F. O. Thomas and H. C. Chu. An experimental investigation of the transition of the planar jet: Subhar-
monic suppression and upstream feedback. Phys. Fluids, 1(9):15661587, 1989. (Three citations in
pages 24, 28, and 123.)
F. O. Thomas and K. M. K. Prakash. An investigation of the natural transition of an untuned planar jet.
Phys. Fluids A, 3(1):90105, 1991. (Three citations in pages 25, 28, and 123.)
TOP 500. TOP 500, November 2010. URL http://top500.org. (Two citations in pages 55 and 56.)
A. A. Townsend. Local isotropy in the turbulent wake of a cylinder. Australian Jour. Sci. Res., ser. A, 1
(2):161174, 1948. (One citation in page 29.)
A. A. Townsend. The mechanism of entrainment in free turbulent ows. J. Fluid Mech., 26:689715,
1966. (Four citations in pages 29, 34, 79, and 85.)
A. A. Townsend. The Structure of Turbulent Shear Flow. Cambridge University Press, Cambridge, 1976.
(Three citations in pages 29, 33, and 104.)
G. Urbin and O. Mtais. Large-eddy simulations of three-dimensional spatially-developing round jets.
In P. R. V. J. P. Chollet and L. Kleiser, editors, Direct and large-eddy simulations II, pages 539542.
Kluwer Academic Publishers, 1997. (One citation in page 27.)
A. Vincent and M. Meneguzzi. The spatial structure and statistical properties of homogeneous turbu-
lence. J. Fluid Mech., 225:120, 1991. (One citation in page 33.)
J. Westerweel, C. Fukushima, J. M. Pedersen, and J. C. R. Hunt. Mechanics of the turbulent-nonturbulent
interface of a jet. Phys. Review Lett., 95:174501, 2005. (Eight citations in pages 29, 31, 32, 75, 80,
83, 87, and 104.)
J. Westerweel, C. Fukushima, J. M. Pedersen, and J. C. R. Hunt. Momentum and scalar transport at the
turbulent/non-turbulent interface of a jet. J. Fluid Mech., 631:199230, 2009. (Nine citations in pages
31, 32, 75, 80, 81, 83, 85, 87, and 104.)
J. Willianson. Low-storage Runge-Kutta schemes. J. Comp. Phys., 35:4856, 1980. (One citation in
page 47.)
I. Wygnanski, M. Sokolov, and D. Friedman. On a turbulent spot in a laminar boundary layer. J. Fluid
140 BIBLIOGRAPHY
A. J. Yule. Large-scale structure in the mixing layer of a round jet. J. Fluid Mech., 89:413432, 1978.
K. B. M. Q. Zaman and A. K. M. F. Hussain. Vortex pairing in a circular jet under controlled excitation.
part 1. general jet response. J. Fluid Mech., 101:449491, 1980. (One citation in page 27.)
J. Zhou, R. J. Adrian, S. Balachandar, and T. M. Kendall. Mechanisms for generating coherent packets
of hairpins vortices in channel ow. J. Fluid Mech., 387:353396, 1999. (One citation in page 19.)

The Dynamics of Coherent Vortices Near The Turbulent/non-Turbulent Interface Analysed by Direct Numerical Simulations

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

The Dynamics of Coherent Vortices Near The Turbulent/non-Turbulent Interface Analysed by Direct Numerical Simulations

Uploaded by

Copyright:

Available Formats

The dynamics of coherent vortices near the turbulent/non-turbulent interface

analysed by direct numerical simulations

, who silently kicked me forward and gave me the last breath

Status not recognized by Republic or Church.

Taylor microscale Reynolds number,Re

turbulent/non-turbulent interface vorticity jump thickness

that describe the ow behaviour of Newtonian uids under the

, which represents the average velocity uctuations

these assume the form:

In isotropic turbulence the ow properties do not exhibit any preferred direction.

. Also in eq. (2.15), represents the characteristic size of eddies at

their characteristic time scale. This leads to the relations

are positive threshold values.

the smallest. A vortex is then dened as a

with the criterion in eq. (2.40)

The authors use the terminology strong vortex instead of at sheets.

)) the criterion uses

(x) is the undisturbed velocity far away from the jet, u

, another relations can

= 0.0165 (see Michalke (1965)).

is roughly equal to the size of the largest coherent

) is given by eq. (2.65):

, such as in a Burgers vortex

/ and the radius

, being N the number of collocation points in each

needed to obtain the solution

Running time of the simulation for the user.

= 1200 in an HIT simula-

, the number of DNS and LES simula-

. The parallel computational problems faced when solving the channel

is the Reynolds number based on the friction velocity u

and the channel half-width. The friction velocity is dened as

1 is the imaginary unit, L

from vector f , can be thus described:

is much cheaper than computing the inversion of the full

) has a quite high cost - proportional to n

does not satisfy incompressibility ( u

, obtained in eq. (3.38), is then corrected by

. Those in between these levels are considered as

is dened by eq. (5.1):

and can be obtained from the histograms in Fig. 7.1

at similar Reynolds numbers.

. Specically, the mean velocity prole is given by,

/ 120 across the jet shear layer.

See sec. (3.6).

20. For the present

, in agreement with other experimental and numerical works (Westerweel

/ 12. Figure 6.4 (b) shows the conditional prole for l

is computed for each one of

, simple dimensional reasoning yields

; (b) Conditional prole of l

depends only on the

(, S), where S is the strain

is the uctuating velocity eld and L

/. Therefore the associated length scale of the

(equation 6.9); (c) Instantaneous prole of vorticity norm and

= 111 and the

roughly constant inside the shear layer

< 1, and also, approximately, for the intense vorticity e.g., [

= /, and maximum azimuthal velocity u

, respectively, where is the circulation,

= 111 and 120, respectively which is well within

obtained in the literature varies between 0.50 < u