Professional Documents
Culture Documents
[ 141 J.
[ 151
[ 161
[ 171
[ 181
AND
I. INTROD.UCTION
OMPUTER communicationnetworksare
a vitalpart of
many
industrial,
governmental,
financial,
and
service
institutions.Theysupportanincreasingvarietyof
services.
Backbone networks today can contain from
a few nodes to
hundreds of nodeswithmanyhostssupportingthousands
of users.Most of thecommercially available networksand
network architectures such
as DATAPAC [ 51 , [ 28 1 , SNA
[ 2 2 ] , [ 2 5 ] , NPDN [ 3 4 ] , TRAIVSPAC [81, TELENET [351,
andTYMNET [ 3 0 ] , [ 3 6 ] have adopted a staticorsemidynamic routing method in which routes are defined at system
generation or atsession initiation for each pairof communicat-
SIDNEY L. HANTLER
n
L
R
6,l
S,
Qz
f r (or {),
1IP
Xr
The
index
set
of the
communicating
source/
destination pairs (commodities) in the network.
Typically, this is a subset of the node pairs.
Theindexset
of linksinthenetwork.
The
index
set
of candidate
routes.
The
set
of
candidate
routes
can
be
provided
by
users,
generated b y a route generation algorithm or a
combination of thesetechniques.
A route is
characterized by the ordered set
of links (from
source to destination) in the route.
An indicator
function
which
is one if link 1 is
used in route r and is zero otherwise.
The index set of candidate routes for commodity
p . We assume that if p # 4 then S, n S, = qj.
Thecapacityoflink
I inbitspersecond.
The message arrival rate of the unique commodity p associated with router , where r E S,.
The
mean
of the
exponential
distribution
from
which the lengthsof the messages are drawn.
decision
A
variable
which
is one if route r is
selected for message routing and zero otherwise.
Using theabovenotationweconcludethatthetotalbit
1 equals & . ~ f r h r l x r / p From
.
this,we
arrival rateonlink
observe that the expected network delayis
ZIP = min
subject t o
I156
TRANSACTIONS
COMMUNICATIONS,
IEEEON
subject to
*,=I
rESp
x r = 0 or 1
VlL
V r E R.
<
<
111. PROBLEMREFORMULATION
In thissection we transformtherouteselectionproblem
formulationintoanequivalentformulationwhich
is better
suitedfora
Lagrangean relaxationprocedure.Letting
fi be
the utilization of link I (i.e.,theproportionofthe
links
capacity consumed by the actual message flow), the objective
function can be rewritten as
where
xr=l
x r = 0 or 1
VrER
IV. LAGRANGEANRELAXATION
By multiplying the constraints, in (3) by a vector of rionpositiveLagrangemultipliers,
X I , I E L,andaddingthem
to the objective function, we obtain the Lagrangean relaxation
of problem IP:
L(h) = min
subject to
O<fr<l
VlEL
and
xr= 1
rESp
and
x, = 0 or 1
VrER.
V. THE SUBGRADIENT
OPTIMIZATION
PROCEDURE
and
or1=
rESp
PQI
r
V r E R andIEL.
1157
motivationforthedecisiontoconcentrateonsubgradient
optimization techniques.
Let xr(h), f l ( X ) be an optimal solution to the
Lagrangean
problem for a fixed vector h. A subgradient is the vector with
coordinates
5) Go t o .2.
Extensivecomputationalexperiencewiththisprocedure
has shown that itis very stable and converges in a few hundred
iterationstoasolutionwhich
is veryclose t o L(h*). The
procedure hasbeen
foundto
beinsensitive
totheinitial
multipliervaluesand
totheinitialoverestimate.However,
70) = f O ) xr(h)arl
'J L.
:,.&he rate of convergence andquality
of boundgenerated
rER
'hepends to a large extent on the limits set on the improvement
Poljack [29 1 has shown that when the multipliers are updated
and iteration counters. Low settings terminate the procedure
using the iterative formula
before it reaches a good solution, while high settings consume
X;+' = hkI + t k Y f
( 5 ) excessive computing resources. Good settings of these parameters require careful balancing between these two considerathen
converges to h* provided that tk converges t o 0 and tions, which seems t o be obtained most readily by experimenztk diverges. Held and Karp [ 2 3 ] and later Held e t al. [24 J tation.
have suggested substituting
1
'
.
x:
VI. SOLVING
THE LAGRANGEAN PROBLEM
tk = 6p
ZIP
0yk
m k )
112
The computational efficiency of the subgradient optimization procedure depends on our ability to solve efficiently the
Lagrangean problemwhich is generatedinstep2)b)
of the
subgradient optimization procedure. Fortunately, for a fixed
setofmultipliers,the
Lagrangean problem is separable into
subproblems which are readilysolved.
The Lagrangean problem can be rewritten as
They have shown that if 0 < 6, < 2-and ZIP < L(h*) the
sequence in (5) converges to h*. With ZIP L(h*) this iterative computation of t k has been used in most successful applications of subgradient optimization procedures.
Thesteps involvedin thesubgradientoptimizationprocedure are as follows.
L(h) = min
{&I+
X,fII
1) Initialization:
a ) Using aheuristic,computeanoverestimate
ZIP of
L(X*) or set ZIP to an arbitrarily large value.
b) Select an initial set of multipliers ho and set k , the
+
IEL
1
0, theimprovementcounter
to 0, X*
iterationcounter,to
t o ho, the current best value of L ( h ) t o 0, and 6 , a parameter
with no change in the set of constraints.
for adjusting the stepsize,
to 6 0 (an arbitrary positiveinitial
Since therearenocouplingconstraintsbetweenthe
fr
value, e.g., 2).
and the x, variables, L(X) can be written as L ( h ) = Ll(A)
2) Solving the Lagrangean problem:
L,(X) where
a ) Incrementtheimprovementanditerationcounters
by 1 .
Subproblem 1
b) Solve the Lagrangean problem using X k as the Lagrange
t
1
multipliers. Thus, obtainL ( h k ) and xk ,fk.
3) Testing and updating parameters:
a ) If L ( h k ) isgreaterthanthecurrentbest
value of
L ( h ) thenreplacethecurrentbest
value of L ( h ) by L ( h k )
andset h* t o hk. Also resettheimprovementcountertosubject
to
>
[z
Z X r { X(-
11
-1.
b) If xk is feasible for
problem
IP,
compute
the
value
0 rfl 1
of its associated objective function for that problem. I
f this
valueis less thanthecurrent
value of Zip, thenset ZIP to and
this
Subproblem 2
C )If the improvement counter has reached a prespecified upper limit, then set 6 to 6/2, X k to h*, the improvement
counter to 0, and go to 2)a).
counter
iteration
d ) If the
mespecified
has exceeded
a
limit 0; if 6 is less than a prespecified limit, or &if t i is less
thana prespecified limit,or if (ZIp - L(X*))/L(h*)is less
than a prespecified error tolerance, then stop.
4) Updating the multipliers: Compute a new subgradient
subject t o
hf+'=min(-l,hf+tkyf)
+ 1.
xr=l
Y P E n
xI=oor 1
'JrER.
LOk)
Subproblem 1 canbeseparated
into J LI subproblems,
one for each link, where the subproblem associated with.the
kth link is
Set k t o k
rESp
'J I E L
'JZEL.
I158
TRANSACTIONS
IEEE
~5
1159
InARPA(Fig.
l),there
are 420 communicatingnode
pairs with two terminals
in each node, each of which causes
two messages persecondoftraffic
t o be generated in each
direction along the chosen route.
In OCT (Fig. 2), there are
650communicatingnode
pairs withoneterminal
in each
nodegeneratingone
messageper secondin each direction.
In RING(Fig. 3), thereare 992 communicatingnode pairs
with one terminalineachnodegeneratingone
message per
second in each direction.
USA (Fig. 4) has 26nodes,each of which contains two
650 communicating pairs witheach
terminals.Thereare
terminalgeneratingtwomessagespersecondineachdirection.
Table I summarizes the results obtained for these networks
with a variety of routes and traffic data. The mean
message
length is measured inbits,theupperandlowerboundsin
milliseconds, andthenumber
of routeslpairdenotesthe
number of candidateroutes allowed foreachcommunicating pair. Some origin-destination pairs admit fewer candidate
routes than others.
In allcases, the system converged without difficulty to a
feasible solution that was near the lower bound or discovered
thatno
feasible solutionexists.Thecomputingtimesand
numbers of iterationsrequired were shortenoughto
allow
for the use of thesystem as aninteractivetool.Depending
on the number of routes involved, each minor iteration consumedbetween 20 and 100 msofvirtual
CPU time on an
IBM 308 1 . Elapsed time between major iterationswas between
1 and 10 s with 300400 VM users contending for resources.
Significantgapsexist
betweenthelowerboundandthe
bestfeasible
solutiononlywhennetworkscontainnearly
saturated links. In practice,this will not be a problem since
network designers will rarelydesign
networksintended t o
operatenearsaturation.
Analysis
of
such cases indicates
that relatively few linksaretypically
bottlenecks. Increasing
thecapacity
oftheselinkseliminates
thebottlenecksand
significantly
increases
the
predicted
performance
of
such
networks.
In some cases,increasing the number of candidate routes
allowsalternativeroutes
t o be chosen,whicheliminatesthe
bottlenecks.Theexpectedperformance
of thenetwork is
sensitive to the number
of routesinthecandidatesetonly
whenthenumber
of routespercommodity
is oneortwo.
When the number of candidate routes per commodity is three
ormore,theoptimalexpecteddelayremainsalmostunchanged.
ARPA
21 NODES
26 LINKS
-50KbitsIsec
OCT
26 NODES
30 LINKS
5OKbitsIsec
Fig. 2.
RING
32 NODES
60 LINKS
IX. CONCLUSIONS
We have used a Lagrangean relaxation procedure to obtain
lower bounds and feasible solutions t o a nonlinear combinatorialoptimizationproblem
arisingin minimizingthemean
delay in networks with given routes and traffic in which delay
is modeled as a network of nonpriority M/M/l queues. Computational experience indicates that this procedure is efficient
and effective in most instances, with duality gaps arising only
in networks near their saturation point.
While subgradientoptimization
seems t o work wellin
practice to solve the relaxed problem, we have no theoretical
results concerning the rate of convergence. Such results would
be desirable, as would computational experience with alternative techniq.ues for solving this nondifferentiable optimization
problem.
We believe thatourtechniquescanbeextendedtonetworks in which delay is modeled as a network of nonpreemptive priority queues. There is a choice of objective functions
in this case, including the possibility of optimizing a weighted
average of the priority delays, and the possibility of optimizing
the high priority delay while ensuring that the lower priority
traffic can be carried. We are also interested in the extent t o
Fig. 3.
-35Kbitskec
___ 33Kbitshec
......... 30Kbitshec
41 LINKS
Fig. 4.
1160
TABLE I
SUMMARY OF COMPUTATIONAL RESULTS ONTHE NETWORKS
IN FIGS. 1-4
T
Generate
AKPA
ARPA
ARPA
15
15
2809
2809
2809
420
420
420
210
210
ARPA
I5
2
2
2
ARPA
ARPA
ARPA
OCT
OCT
OCT
OCT
OCT
OCT
OCT
ARPA
ARPA
210
325
598
598
Percent
Length
50
60
65
50
60
65
50
60
7.420
7.575
11.946
15.790
7.432
11.985
15.913
8.040
16.070
0.733
0.715
16.162
7.446
12.005
15.954
8.040
16.070
2.306
0.189
0.167
0.254
0.000
0.000
0.000
0.000
0.220
0.390
0.398
1.958
0.996
2.090
1195
1195
200
220
8
16
2294
2886
220
220
496
100
496
496
77.256
330.750
44.443
49.874
59.563
330.750
45.100
51.232
61.608
1481
1984
1984
I80
183
183
190
200
200
200
210
55.556
53.322
62.831
56.817
54.805
64.006
2762
2762
32
35
6.418
7.288
6.502
7.712
RING
2
2
2
3
4
4
USA
USA
BC
Lower
RING
RING
Message
598
598
RING
TI
65
I80
200
210
21 5
220
2
2
2
2
OCT
OCT
RING
RING
RING
RING
RING
Mean
992
992
992
206.980
102.200
81.276
108.035
128.250
155.890
80.912
154.970
206.980
102.200
81.456
108.459
128.765
159.010
81.720
158.280
150.780
153.370
152.070
13.929
13.929
77.256
1.690
0.847
0.000
12.032
153.370
0.000
0.000
1.456
2.650
3.319
2.220
2.705
I . 836
1.290
5.496
whichourmethodscan
be extendedtonetworkswith
different models of delay, e.g., MJMJmqueues.
The ideas that were presented in this paper can be applied
to related network design pToblems. Work is now in progress
[ 171 on developing modelsandoptimizationtechniques
forthesimultaneousroutingandcapacityassignmentproblemsincomputercommunicationnetworks.Thesemodels
incorporatelinks,traffic,anddelaycostdatatoselectlink
capacities and route the traffic,
REFERENCES
P.Afentakisand
B.Gavish, Optimallot-sizingalgorithms
for
complexproductstructures,Grad.SchoolManagement,Univ.
Rochester, Rochester, NY, Working Paper QM-8318, 1983.
V. Ahuja, Routing and flow control in systems network architecture, IBM Syst. J . , vol. 18, pp. 298-314, 1979.
D. P. Bertsekas, A class of optimal routing algorithms for communications networks, in Proc. 1980 Int. Conf. Circuits Comput.,
Atlanta, GA, Nov. 1980.
D. G . Cantor and M. Gerla, Optimal routing in a packet switched
computer network, IEEE Trans. Comput., vol. C-23, pp. 10621069,1974.
P.M.Cashin,Datapac
networkprotocols, in Proc.3rdInt.
Conf. Comput. Commun., Toronto, Ont., Canada, 1976, pp. 15&
155.
P. J . Courtois and P. Semal, An algorithm for the optimizationof
nonbifurcated flows in computer communication networks, Perf o r m . E v a l . , vol. 1, pp. 139-152, 1981.
A. Danet, R. Despres, A. Le Rest, G . Pichon, and S . Ritzenthaler,
The Frenchpublicpacketswitchingservice:TheTRANSPAC
network, in Proc. 3rd Int. Comput. Commun. Conf., 1976, pp.
25 1-260.
R. Despres andG . Pichon, The TRANSPAC network status report
IEEE TRANSACTIONS
ON
1161
Absrmcr-A
combined
random/reservation
multiple
access
(CRRMA) scheme for packet-switched communication over a global
beam satellite with on-board processing is proposed and analyzed.
Channel time is divided into contiguous
slots; each slot contains N
minislotsfortransmission
of requestpacketsand
N minislots for
data. With N substantially smaller than the number of earth stations,
collisions will occur in requestpackettransmissions.
Two channel
access algorithms for the CRRMA model are proposed: uncontrolled
channel access (UCA) and controlled channel access (CCA). UCA is
simpler but has an inherent stability problem particularly when the
number of minislots N is small. TheCCAalgorithmrestrictsthe
transmission of request packets for new arrivals to take place only
when the slot is in the FREE state. With N = 3, the CCA algorithm
exhibits good delay-throughput characteristics. As N increases, the
UCA algorithm offers stable operation. For N 2 5 the simpler UCA
algorithm is preferred over CCA.
I . INTRODUCTION
largenumberofpapersonmultipleaccesstechniques
ranging from pure random access to totally coordinated
dynamic assignment access have appeared in the literature.
As
they are too numerouslistto them all, only those whicharepertinent to the present paper are listed in the References.
A primary concern in packet switched data communication
over a broadcast channel is the problem
of organizing the access
of many users which share the common channel. Twoimportant
performancemeasuresinselectingasuitablemultipleaccess
scheme are 1) maximum attainable channel utilization and 2)
meanpacketdelay.Theseperformancemeasuresdependon
the system parameters such as the size of the user population,
the burstiness of data traffic, and the propagation delay. In
this paper we consider a very large user population with bursty
trafficandachannelwithalongpropagationdelay,
e.g., a
satellite channel.
The assumption of a large population of bursty users excludes any consideration of fixed assignment schemes such as
frequency division multiple access (FDMA) and synchronous
timedivisionmultipleaccess(STDMA)whichpermanently
assign a portion of the channel capacity to an individual user,