EnergyefficientdatatransmissionusingApproximate

DynamicProgramminginMobileCloudComputing
Authors Name: K MANIKANDAN
CSE DEPARTMENT
R.M.K ENGINEERING COLLEGE
CHENNAI, TAMILNADU
kmanikandanslm@yahoo.com

Authors Name R.RENUGA DEVI M.E.,(Ph.D)

CSE DEPARTMENT
R.M.K ENGINEERING COLLEGE
CHENNAI, TAMILNADU

AbstractMobile cloud computing (MCC)

converges mobile computing and Cloud
computing for augmenting resource-poor
mobile devices to run heavier applications.
With the increasing ubiquity of mobile
devices, MCC exhibits vast application potential in various areas. Energy-efficient data
transmission is a key issue in MCC due to
energy-poverty of mobile devices. In this
letter, we address the issue of energy-efficient
link selection and data transmission
scheduling for delay-tolerant and dataintensive applications in MCC. We first
formulate the problem as a discrete-time
stochastic dynamic program (SDP) that aims
to optimize both system throughput and
energy consumption. To solve the formulated
SDP, we then propose a scalable approximate
dynamic programming (ADP) algorithm that
does not require the statistics of exogenous
stochastic information (e.g., data arrival).
Simulation studies show that the proposed
ADP algorithm can reduce the average
energy consumed for delivering a packet by a
maximum of over 40 percent compared to
alternative minimum-delay and SALSA
policies.
Index TermsEnergy consumption,
transmission scheduling, link selection,
approximate dynamic programming, mobile
cloud computing.
I. INTRODUCTION
Mobile cloud computing is emerging as a new
computing paradigm that aims to augment resourcepoor mobile devices taking advantage
of the abundant resources hosted by Clouds. It makes
running computing applications feasible on resourceconstrained mobile devices, such

as augmented reality and remote healthcare.

However, to fully unleash the application potential
of MCC, energy-efficient data transmission between
mobile devices and Clouds is essential due to energypoverty of mobile devices. As per the 2012 U.S.
Wireless Smartphone and Traditional Mobile Phone
Satisfaction Studies, satisfaction with battery life is
the least satisfying attribute of smartphones, scoring
only 6.7 out of 10, down from 6.9 in 2011.
Fortunately, many mobile applications are naturally
delay-tolerant, e.g., social network service (SNS),
offering mobile users energy-saving opportunities
that can be seized by flexible link selection and data
transmission scheduling. Here a link refers to a
usable Internet connection that is provided through an
ISPs cellular network, e.g., 3G, 4G LTE, or WiFi
access points (APs). For example, when receiving
push notifications of SNS live feeds on mobile, one
may defer the checking of the updated contents until
a high-quality cellular connection or a WiFi AP is
available, considering: 1) the WiFi interface is, in
general, more energy-efficient than the cellular
interface ; and 2) data transmission in good
connection consumes much less energy than in bad
connection .
In this letter, we address the issue of energyefficient uplink and downlink data transmission for
delay-tolerant and data-intensive applications in
MCC. This is challenging because the statistics of
exogenous stochastic information, e.g., data arrival,
link availability, are usually absent in practice. There
have been some studies devoted to energy-efficient
data trans-fer of mobile devices by means of dynamic
link selection and data transmission scheduling. For
example, in, the authors applied these techniques to
studying the energy-delay tradeoffs of either uplink
or downlink data transmission between mobile
devices and infrastructures. Our work differs from
them substantially with respect to the following
critical aspects:

We formulate the link selection and transmission

schedul-ing problem as a discrete-time SDP,
which considers both uplink and downlink data
transmission, and jointly optimizes system
throughput and energy consumption.
We propose a scalable ADP algorithm that does
not require the statistics of exogenous stochastic
information, and solves the curse of
dimensionality
that
generally
plagues
stochastic dynamic programs.
We demonstrate the effectiveness of the ADP
algorithm as well as its superior energyefficiency over alternative policies through
simulation studies.

Fig1.EXAMPLE

Recent smartphones have multiple wireless

interfaces 3G/EDGE (Enhanced GPRS) and WiFi
that can be used for data transfer. These two radios
have different characteristics. First, their nominal
data rates differ significantly (from hundreds of
Kbps for EDGE, to a few Mbps for 3G, to ten or
more Mbps for WiFi). The achievable data rates for
these radios depends upon the environ-ment, can
vary widely, and are sometimes far less than the
nominal values. Second, their energy-efficiency also
differs by more than an order of magnitude . While
the power consumption on the two kinds of radios
can be comparable, the energy usage for transmitting a fixed amount of data can differ an order of
magnitude or more because the achievable data
rates on these interfaces differ significantly. Finally,
the availability characteristics of these two kinds of
networks can vary significantly. At least as of this
writing, the penetration of some form of cellular
availability (EDGE or 3G) is significantly higher
than WiFi, on average. A similar observa-tion has

been made in where the authors report 99% and

46% experienced availability, respectively, in their
traces for EDGE and WiFi. Thus, uploading or
downloading large data items using WiFi can be
more energy-efficient than using the cellular radio,
but WiFi may not always be available.
Contributions. In this paper, we present a principled
approach for designing an online algorithm for this
energy-delay tradeoff using the Lyapunov
optimization framework. We formulat the link
selection problem as an optimization formulation
which minimizes the total energy expenditure subject
to keeping the average queue length finite. The
Lyapunov optimization framework enables us to
design a control algorithm, called SALSA (Stable and
Adaptive Link Selection Algorithm), that is
guaranteed to achieve near-optimal power
consumption while keeping the average queue finite .
Specifically, we show that, in theory, SALSA can
achieve power consumption arbitrarily close to the
optimal. To our knowledge, prior work has not
explored this link selection problem, and our use of
the Lyapunov framework for solving this problem is
also novel.Our second contribution is an exploration
of two issues that arise in the practical
implementation of SALSA. First, although control
algorithms based on the Lyapunov framework have a
single parameter V, the theory does not give any
guidance on how to set that videos. Figure 1 presents
a screenshot of the systems Web interface, showing
some user-generated video-clips from our users.Our
users report that battery lifetime is a critical usability
issue,and video uploads use a significant fraction of
the energy in our system. This paper explores robust
methods for reducing this cost.Recent smartphones
havemultiple wireless interface-3G/EDGE (Enhanced
GPRS) and WiFi that can be used for data
transfer.These two radios have widely different
characteristics. First, their nominal data rates differ
significantly (from hundreds of Kbps for EDGE, to a
few Mbps for 3G, to ten or more Mbps for WiFi).
The achievable data rates for these radios depends
upon the environment, can vary widely, and are
sometimes far less than the nominal values. Second,
their energy-efficiency also differs by more than an
order of magnitude. While the power consumption on
the two kinds of radios can be comparable, the energy
usage for transmitting a fixed amount of data can
differ an order of magnitude or more because the
achievable data rates on these interfaces differ
significantly. Finally, the availability characteristics
of these two kinds of networks can vary significantly.
At least as of this writing, the penetration of some
form of cellular availability (EDGE or 3G) is

significantly higher than WiFi, on average. A similar

observationhas been made in [22] where the authors
report 99% and 46%experienced availability,
respectively, in their traces for EDGE and WiFi.
Thus, uploading or downloading large data items
usingWiFi can be more energy-efficient than using
the cellular radio, butWiFi may not always be
available.

II.Performance against Baseline Algorithms.

We first compare SALSA against the two baseline
algorithms, MINIMUM-DELAY and WIFI ONLY
plots the performance of each of these algorithms on
the E-D plane. For SALSA, we use an a of 0.2: we
later explore the performance of SALSA across a
range of a values. Each point on the E-D plane
corresponds to one simulation run. One way of
characterizing the overall performance of the
algorithm is to understand the shape of its envelope
the convex hull of all the points on the E-D plane.
For the class of algorithms that make an energy delay
trade-off, what characterizes a good envelope?
Intuitively, a good algorithm should be capable of
achieving a good balance between energy and
delay: neither the delay per byte, nor the energy per
byte should be too large. In other words, the points on
the E-D plane should be clustered around the origin,
and the envelope should be compact. We use this
intuition to compare different algorithms throughout
this section.

and samples the exogenous stochastic processes t.

Unlike classic backward dynamic programming, it
then steps forward through time t and relies on an
approximate value function V a(sat) parameterized
by for decision-making. The statistical
information of t is not required because it has been
captured in the sample observation v(st), which is, in
turn, used to update the parameter vector . As the
iteration moves forward, will converge to a certain
value. The learning process then terminates, and the
obtained policy can be used for energyefficient
link selection and transmission scheduling in MCC.
The time complexity of Algorithm 1 is O(4LT) where
L _ maxt{|Lt|} is the maximum number of usable
links.Since the algorithm maintains only a single
parameter vector for iterative updating, it does not
impose a heavy burden on computation or memory
when implemented on mobiles.
Algorithm An ADP algorithm for solving problem P
Input: An iteration number T , basis functions (sat).
Output: .
1: Initialization: 0, s0 = (q0 0, 0).
2: for t = 0 to T 1 step 1 do
3: State observation: observe the current system
state t;
4: Link selection and transmission scheduling:
solve the problem (14), and let [lt ; at ] be the
decision that maximizes v(st);
5: if t > 0 then
6: Parameter updating: update end if
8: State transition: update sat by taking [lt ; at ]
compute V a(sat); sample t+1 and update st+1.
9: end for
10: return .

IV. SIMULATION RESULTS AND

CONCLUDED REMARKS
A. Simulation Setup

Fig.2.SALSA ENEREGY SAVINGS

III. An Approximate Dynamic Programming

Algorithm
Algorithm 1 presents the ADP algorithm for solving
problem P. At the beginning of each slot t, the
algorithm first observes the system queue backlog qt

We assume that the uplink and downlink data arrivals

follow Poisson processes with E{t} = (2, 9) packets
per second. Without loss of generality, we assume
that all packets are of size 50 KB. The buffer size
limit q is set to (100, 500) packets. In each slot t, the
availability of the cellular/WiFi interface is modeled
as a two-state Markov chain while the uplink and
downlink bandwidth is randomly and uniformly
distributed, as parameterized in Tab I. Tab. II presents
the parameter settings of the energy model introduced
in Sec.II. It is worth highlighting that the stochastic
processes.

Our simulation involves two phases: learning

iterations where at each iteration we sample a
realization of t, solve the problem and update ;
and testing iterations where we fix the obtained
and rerun Algorithm 1 on a different sample path
(t)T1 t=0 . We use T = 2000 and = 0.99 such that
T < 109 is sufficiently small to approximate the
infinite horizon problem . Each iteration lasts = 30
seconds.

V Simulation Results
First, verification of the convergence of the ADP
Algorithm. plots the rate of convergence of the
parameter vector under three different stepsize
recipes with the corresponding parameter in brackets
It shows that converges around the point (4400,
4800)T as the iteration moves forward. In particular,
we observe apparent convergence of around 500
iterations. This is a fair rate of convergence
considering the complexity of the multidimensional
exogenous stochastic information t.Second,
comparison with alternative policies. We simulate
two alternative policies for comparison: SALSA(V) as
proposed in where V is a control knob, and the
minimum-delay policy therein. For ease of
comparison, we define a performance metric Ep,
referred to as the average energy consumption per
packet, that evaluates both system throughput and
energy consumption.

Fig.3.PERFORMANCE OD ADP ALGORITHM

VI Implementation Issues
The ADP algorithm was tested through Monte Carlo
simulation in an offline mode. In practice, to speed
up the rate of convergence, mobile users can first use
collected workload and link trace to train the
parameter vector offline for a period of time, and
then put it online for on-the-fly training. For mobile
applications with stringent delay bounds, offline
training is particularly important for performance
guarantees. Nonetheless, it may not be necessary for
delay-tolerant applications with longer thresholds.
Our future work will focus on implementing the ADP
algorithm on a smartphone chip, and testing its
performance on real devices.