See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/220640835

Topological design of interconnected LAN/MAN networks

Article in IEEE Journal on Selected Areas in Communications · October 1993

DOI: 10.1109/49.245906 · Source: DBLP

CITATIONS READS

59 537

2 authors:

Cem Ersoy Shivendra S. Panwar

Bogazici University Polytechnic Institute of New York University
266 PUBLICATIONS 10,307 CITATIONS 325 PUBLICATIONS 8,300 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Cem Ersoy on 31 May 2014.

The user has requested enhancement of the downloaded file.

1172 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 11, NO. 8, OCTOBER 1993

Topological Design of
Interconnected LAN/MAN Networks
Cem Ersoy, Member, IEEE, and Shivendra S. Panwar, Member, IEEE

Abstract- This paper describes a methodology for designing

interconnected LAN/MAN networks with the objective of mini-
mizing the average network delay. We first consider IEEE 802

L**
standard LAN’s interconnected by transparent bridges. These
bridges are required to form a spanning tree topology. We
propose a simulated annealing-based algorithm for designing
minimum delay spanning tree topologies. In order to measure
the quality of the solutions, we find a lower bound for the LAN
average network delay. We extend the algorithm to design the
0 LAN-LAN Bridge
overall LAN/MAN topology consisting of a MAN or high-speed
data service interconnecting several clusters of bridged LAN’s. la] LAN-MAN Bridge
Comparison with the lower bound and several other goodness
measures show that the solutions are not very far from the global Fig. 1. An example LAN/MAN topology.
minimum.
spanning tree bridges cannot be used to allow all source
I. INTRODUCTION destination pairs to use the best path that exists between them

L OCAL area networks (LAN’s) have become an indis- in the physical topology. However, some degree of flexibility
pensable part of the modem working environment. Wide- is available since network management can be used to force the
spread applications using LAN’s and a continuously growing shape of this deterministic logical topology to be any spanning
number of users have created a necessity for the interconnec- tree. The performance of bridged LAN’s has been studied
tion of LAN’s. Bridges, routers, and gateways are different in [3]-[5]. The problem of determining which gateways to
interconnection devices used to overcome the geographical use to interconnect existing data networks has been discussed
and capacity limitations of LAN’s. Among these, bridges in [6]. A simulated annealing technique has been used for
are the simplest since they operate at the two lowest layers designing minimum-cost interconnected CSMA/CD LAN’s in
of the Open Systems Interconnection (0%)model, namely [7]. The design of optimally locating bridges and repeaters for
the link and physical layers. Different types of bridges are minimizing the average delay and fast algorithms for special
used for interconnecting various types of LAN’s. We consider cases have been studied in [8].
transparent bridges which are widely used for interconnecting There is a growing interest in the interconnection of ge-
CSMA/CD LAN’s and other IEEE 802 standard LAN’s [l]. ographically separated LAN’s or bridged LAN clusters by a
The major advantage of transparent bridges is that they do not backbone metropolitan area network (MAN). Fig. 1 shows
require the participation of end users in order to interconnect an example interconnected LAN/MAN network with several
LAN’s. These bridges are self-learning and self-configuring. clusters. Each cluster consists of several LAN’s interconnected
They learn and store the location of end users by observing by transparent bridges. The major reason for forming these
source addresses of data packets. They route packets by clusters is the geographical separation of LAN’s. There may
comparing the destination addresses to the table of learned also be network management or security reasons for keeping
addresses. In order to operate properly, they require the logical several LAN’s in a cluster. The overall topology has to be a
topology of bridges and LAN’s to be a spanning tree [2]. spanning tree since this is required by the bridges. Depending
Given an arbitrary physical topology, these bridges configure on the physical distances involved and the capacity required,
themselves to be part of a deterministic active spanning tree different backbone network architectures can be used for
topology. One of the main disadvantages of this is that interconnecting clusters of bridged LAN’s. Fiber Distributed
Data Interface (FDDI), which is a 100 Mb/s network based
Manuscript received June 1992; revised December 1992. This work was
supported by the New York State Science and Technology Foundation’s Center
on a token ring protocol, has already been used for the
for Advanced Technology in Telecommunications, Polytechnic University, interconnection of LAN clusters [9]. There are other promising
Brooklyn, NY and by the NSF under Grants NOR-8909719, NCR-9003006, candidates such as backbone MAN’S interconnecting LAN
and NCR-9115864. This paper was presented in part at the IEEE INFO-
COM’92, Florence, Italy, May 6-8, 1992. clusters. One of them is the IEEE 802.6 MAN standard: Dual
C. Ersoy is now with the Department of Computer Engineering, Bogazici Queue Dual Bus (DQDB), which has typical data rates ranging
University, Istanbul, Turkey. from 50 to 150 Mb/s [lo].
S. S. Panwar is with the Department of Electrical Engineering, Polytechnic
University, Brooklyn, NY. There are also proposals of broadband services such as
IEEE Log Number 9211669. Switched Multimegabit Data Service (SMDS) for intercon-
0733-8716/93$03.00 0 1993 IEEE
ERSOY AND PANWAR: TOPOLOGICAL DESIGN OF INTERCONNECTED LAN/MAN NETWORKS 1173

necting LAN clusters using MAN technology [ 11). SMDS is a

very high-speed connectionless packet-switched service which
is to be offered by the local exchange camers in the early
1990’s. This service is suitable for bridging between IEEE
802 LAN’s since it can handle variable size packets of up to
8192 octets. SMDS will provide services with DS1 (1.5 Mb/s)
and DS3 (45 Mb/s) rates. The DS3 rate will have access class
rates of 4, 10, 16, and 34 Mb/s. The SMDS Interface Protocol
is aligned with the IEEE 802.6 MAN standard in order to be
suitable for LAN interconnection, since this is expected to be
an important application of SMDS.
The problems related to using B-ISDN (Broadband Inte-
grated Services Digital Network) for the interconnection of
LAN’s have been addressed by Mongiovi et al. in [12]. Since Fig. 2. An example of a bridged LAN network.
the proposed B-ISDN standard, Asynchronous Transfer Mode
(ATM) [ 131, is connection oriented and divides the information
into fixed-size cells, it is not directly compatible with the IEEE interconnected by local transparent bridges. In order to avoid
802 LAN protocols. ISDN Frame Relay Service has also been looping of packets, these bridges require the active topology to
suggested for LAN interconnection [ 141. be a spanning tree [2]. We will describe how individual cluster
In this paper, we address the problem of overall LAN/MAN topologies can be combined to form an overall LAN/MAN
topology design so that the average network delay is min- topology in Section 111.
imized and the maximum end-to-end delay of the overall
topology is below a threshold. Topological design of the
overall LAN/MAN topology as a whole is very complex and A . Definition and Formulation of the Problem
difficult since the total number of LAN’s is large and many An example physical topology for a group of bridged LAN’s
constraints are involved. In order to reduce the complexity, we
is shown in Fig. 2. One of the possible active spanning
define and formulate the problem so that it can be decomposedtrees is shown with bold lines. During normal operation,
into smaller problems for each cluster. We show that subprob-only the bridges on the active tree forward packets. Other
lems for each cluster can be solved independently. After the bridges remain idle unless activated to form a new tree in
decomposition, each subproblem involves only the LAN’s in case of failures. Since the logical spanning tree topology is
that cluster and the backbone MAN. Section I1 covers the independent of the underlying physical topology, real costs
design of individual cluster topologies. In Section 11-A, thedo not play an important role in the logical topology design
minimum-delay spanning tree problem for the interconnection problem. Hence, performance measures can be used as the
of LAN’s in a cluster is defined and formulated as a com- objective function for finding the optimum logical topology.
binatorial optimization problem. There is no efficient exact We address the problem of designing minimum-delay spanning
algorithm for solving this problem. We propose a heuristic tree topologies for the interconnection of LAN’s. In this
algorithm based on simulated annealing for finding solutionsproblem, N LAN’s, local traffic on each LAN ( t i i ) ,and the
to the minimum-delay spanning tree problem in Section II- traffic requirements between all source and destination LAN
D. The spanning tree design can then be implemented either pairs ( t s d ) are given. Our goal is finding the spanning tree
by connecting the bridges to the LAN’s appropriately or topology with the minimum average network delay for the
by adjusting network management parameters if the existing given set of requirements. In other words, we have to decide
physical topology can be pruned to the desired tree. In order
where to place ( N - 1) active bridges among all possible LAN
to check the quality of solutions found by the simulated pairs such that they form a spanning tree with the minimum
annealing algorithm, a lower bound for the problem is found delay. In a given network, if the underlying physical topology
in Section 11-C. The algorithm is extended for the design can be pruned to the minimum-delay spanning tree, bridges
of overall LAN/MAN topologies in Section 111. Results of can be set up to choose this tree as the active logical topology
the computational experiments given in Section IV show that by adjusting the parameters of the self-configurationalgorithm
by decomposing the problem and using simulated annealing, 1151.
low-delay LAN/MAN topologies for fairly large networks As in Fig. 2, LAN’s are represented as nodes and bridges
can be found in reasonable running times. The delays of as edges of a graph. Since many existing bridges have two
these topologies are lower than those found by other feasible
ports and we are interested in the shape of the minimum-delay
algorithms. Section V concludes the paper. topology, this representation is sufficient for our purposes.
With slight modification, multiport bridges can be represented
as several edges and a central node. In that case, edges of
11. DESIGNOF MINIMUM DELAYINTERCONNECTED LAN’s the graph will correspond to bridge ports rather than whole
In this section, we will describe the topological design bridges. In this way, both two-port and multiport bridges
problem for bridged LAN’s which belong to the same cluster. can be used for implementing minimum-delay spanning tree
The LAN’s are located geographically close to each other and topologies.
1174 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 1 1 , NO. 8, OCTOBER 1993

The problem of finding the minimum-delay spanning tree Because of the large number of topologies generated, we need
topology is formulated as described: an efficient way of approximating the average network delay;
Problem P: preferably, a closed-form delay model. We consider delays
due to LAN’s and bridges. The topology is modeled as a
network of queues. Different measurement and performance
studies have shown that LAN traffic is bursty and consists of
N N \ batches or trains of packets because of the existing protocols
[20], [21]. In order to account for the burstiness of traffic
and to have a computationally inexpensive measure, we use
subject to M X / M / l queues with batch Poisson arrivals. Although it is
simple, M X / M / 1 queueing model may achieve an acceptable
fi < p i VLANi (2) fit for the busiest periods of the network [21] and give us the
f i j < xijpij VBridgeij (3) relative performance of different topologies. Other closed-form
delay models may also be used in the algorithm. However,
where more complex delay models will increase the running times.
1, if there is a bridge We are given the mean rate for traffic requirements between
decision variables xij =
{ between LAN i and LAN j
0, otherwise.
S = set of all possible spanning trees.
each LAN pair ( t s d ) in terms of batches per second. For a
given spanning tree topology, we can find the mean batch
flow values of LAN’s and bridges A; and A i j , respectively. We
know the capacities of LAN’s, Ci (bitdsecond). We assume
The objective function is the average network delay (T) which exponential distribution for packet lengths with mean 1 and
is a combination of the queueing delays due to LAN’s (E[T;]) a geometrical distribution for the number of packets in a
and bridges (E[Tij]) as described in [16]. E[T;]is a function of batch with mean X. The mean service rate for LAN’s is
the total flow on LAN i (f;) and E[T;j]is a function of the total equal to the LAN capacity divided by the mean packet length
flow on the bridge between LAN i and LAN j (fij). y is the (pi = C;/l packets/second). The mean service rate for each
total input flow to the network. We do not consider propagation direction of the bridges is equal to p ; j (packetshecond). The
delays since LAN’s are geographically close to each other. A expected number of packets ( E [ L ] )in each queue is found
more detailed description of the delay model will be given in using the results in [22, pp. 156-1601. Using Little’s formula
Section 11-B. pi and p i j are the LAN and bridge capacities, ( E [ T ]= E [ L ] / X X )the , expected value of the delay ( E [ T ] )
respectively. Inequalities (2) and (3) are the corresponding for each M X / M / l queue is given as
LAN and bridge capacity constraints. They are also implied in
the objective function, since the delay ( T )becomes infinity if P AX
E[T]= ___ wherep = -. (4)
the flow values exceed the capacities. In order to keep the 4 1 - P) CL
problem simple, we do not consider additional constraints.
After finding the delays caused by LAN’s and bridges using
If needed, other constraints such as communication costs or
(4),the average network delay ( T ) is found by plugging
maximum utilization for LAN’s and bridges can easily be
these delay values, total flow values in terms of packets per
incorporated into the formulation.
second (f = AX), and the total input flow of the network
Problem P is a difficult combinatorial optimization problem.
t s d ) into the objective function of (1).
In fact, a simplified version of this problem, namely the (y = X
capacitated spanning tree problem, which handles queueing
delay implicitly, is NP-complete [17]. There is no efficient C . A Lower Bound on the Average Network Delay
exact algorithm for solving Problem P. The number of possible The average network delay is a combination of delays due
spanning trees for N LAN’s is equal to N N - 2 [18]. This to LAN’s and bridges TL and T B ,respectively:
makes it impractical to determine the minimum-delay topology
by exhaustive search except for very small size problems. l N
T =-xX;XE[T;]
Our experiments showed that the solution space has many Z=l
local minima and greedy local search heuristics get stuck in a
local minimum. In order to avoid this problem, we based our
search heuristic on simulated annealing since this technique
works well with problems with many local minima [19]. For
small problems, we found the global minimum with complete
enumeration. In order to check the quality of the solutions
- TB
for larger problems, we derived a lower bound for Problem P
As shown in ( 3 , the sum of the lower bounds on the delay
which is described in Section 11-C.
due to bridges and LAN’s will give us the lower bound on
T. In order to find a bound on the delay due to bridges ( T B ) ,
B . Delay Model we first find a lower bound on the total flow on all bridges
We find the average network delay for each topology gen- (A;). Then, we find the optimum distribution of this total flow
erated at every iteration of the simulated annealing algorithm. among all bridges such that TB is minimized. Similarly, we
ERSOY AND PANWAR: TOPOLOGICAL DESIGN OF INTERCONNECTED LANMAN NETWORKS 1175

find a lower bound on the total flow on all LAN's (A:). The where w:d is equal to 1 if LAN i is on P a t h s d , zero otherwise.
optimum distribution of this flow among all LAN's such that Therefore,
TL is minimized will give us the lower bound on TL. The
lower bound on T obtained by summing the lower bounds on A T = TLIA 2 (12)
TL and TB enables us to check the quality of the solutions i
found by simulated annealing and other techniques.
1 ) A Lower Bound on the Total Flow on Bridges: The total
flow on all bridges (A): is equal to the sum of all bridge flows.
The spanning tree topology with the minimum A; gives us a
lower bound. s d

Definition 1 : Optimum requirement spanning tree problem

[23]: In this problem, the cost of communication for a span- s d
ning tree is equal to the summation of all the products of a
Since requirement t s d passes through all the LAN's on P a t h s d
traffic requirement ( t s d ) times the number of hops on the path
including LAN s and LAN d , the number of times it will
corresponding to that requirement ( h s d ) (i.e., C, E d t s d . h s d ) .
appear in the summation is equal to one more than the number
The spanning tree which minimizes this cost of communication
of hops on P U t h , d . The local traffic ti; affects only LAN 2.
is called the optimum requirement spanning tree and can be
All local traffic requirements appear in the summation once.
found with an O(N4) algorithm [23].
The minimum value for the total flow on LAN's is
Theorem 1 : The optimum requirement spanning tree de-
equal to the summation of A;,min and all t s d ' S , since only the
fined by Hu in [23] has the minimum A; among all spanning
trees. terms involving A; can be minimized.
Proof: The flow on the bridge between LAN i and LAN 3 ) Optimum Distribution of Total Flows: After obtaining the
j ( B r i d g e i j ) is equal to the sum of all requirements ( t s d ) lower bounds on the total flow on LAN's and bridges, the
which have to pass through B r i d g e i j in order to reach their problem of finding the lower bounds on TL and TB reduces
destination to finding the optimum distribution of these total flows so that
TL and TB are minimized. In order to make the lower bounds
tighter, we also consider some inequality constraints.
s d
Any traffic requirement which has LAN i as its source or
destination has to pass through LAN i. Therefore, the flow on
where y z is equal to one if B r i d g e i j is on Path,d, zero LAN i has to be greater than or equal to the mandatory flow
otherwise. Therefore, on that LAN (Ay):
N
A; = Aij (7)
VBridge,?
k=l

In other words, the flow on LAN z consists of the summation of

its mandatory flow (Ay)and the transit flow on LAN i (ATR):
(9)
A, = A," + A:~, where
AT" >O Va. (17)
and since for a given (s, d ) pair the number of nonzero y s ' s
is equal to the number of bridges on P a t h s d , the number of Finding a lower bound for TL is equivalent to finding the
times t s d appears in the summation is equal to the number optimum distribution of A:,m2, which minimizes TL subject
of hops in P a t h , d . Thus, the total flow on bridges can be to the inequality constraints in (16). This problem is similar to
written as the capacity assignment problem described in [16], but here the
capacities and the total flow are known and individual LAN
flows are to be found. The optimum distribution of TL is a
s d standard optimization problem and, as described in Appendix
A, it can be solved using Lagrangean techniques. In Appendix
which is also equal to the cost of communication for the op- A, we also show by using the Kuhn-Tucker conditions [24]
timum requirement spanning tree problem in [23]. Therefore, that if all LAN capacities are equal, the water-filling type
the cost of the solution to this problem gives the lower bound distribution shown in Fig. 3 minimizes TL.In this distribution,
';,min. 0
2 ) A Lower Bound on the Total Flow on L A N ' s : Similarly, AT" = (A; - A,"')' vz (18)
the total flow on all LAN's (A): is the summation of flow
where A i is chosen so that
on all LAN's. The flow on LAN i ( & ) can be found as

s d
c(~i
N

2=1
- A,")+ = AE,rnzn -
N

2=1
P' (19)
1176

Flows
LAN

and
L
I

(I)+
LAN1
I

LAN2 LAN3 LAN4

Fig. 3. Distribution of LAN flows

denotes the positive part of E , i.e.,

LAN5

Similarly, as described in Appendix A, a lower bound for

the delay due to bridges ( T B )can be obtained by optimally
distributing on 2(N-1) bridge ports so that TB is min-
-
IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 11, NO. 8, OCTOBER 1993

I I
I

*
’
4
New-Tree =
Neighbor of Current-Tree
I

New-Trec feiuible 1

Keep the best tree so far

Stop annealing 1
I
I

imized. We also show that if all bridges have equal capacity, ( Save the best tree so far

a uniform distribution of among all bridge ports gives

a lower bound for TB.This lower bound can be improved by Fig. 4. The flowchart of the simulated annealing algorithm.
considering additional inequality constraints similar to the ones
described for LAN’s. This time the mandatory traffic concept called neighbor trees. Given a spanning tree, we can create a
will be replaced by the 2 ( N - 1) largest traffic requirements. neighbor tree by removing a branch, resulting in two separate
Since each of these traffic requirements has to pass through at subtrees. Adding another branch which will connect the two
least one bridge, they impose inequality constraints on bridge subtrees, but will not create a loop. During the search for a
flows. In order to minimize T B ,each of these requirements is minimum-delay topology, we create a sequence of neighbor
assigned to a separate bridge port. As in the case of LAN’s, trees as described in the flowchart of Fig. 4. We check the
it can be verified using the Kuhn-Tucker conditions that a feasibility of each topology by comparing the flows of LAN’s
water-filling type distribution similar to the one in Fig. 3 is and bridge ports with their capacities. If the topology is
optimum and gives a lower bound for TB. However, unlike feasible, we find the average network delay using (1). During
the LAN case, if the bridge port capacities are not equal, these the local neighborhood search for a minimum-delay spanning
inequality constraints cannot be used for improving the lower tree topology, if the transition from the current topology to a
bound on TB.After finding the lower bound for TL and T B , new topology is a delay-decreasing one, it is always accepted;
the sum of these terms gives the lower bound for the average if the transition is a delay-increasing one, the new topology
network delay T . is accepted as the current solution with a certain probability,
p , which is given by
D. Simulated Annealing Algorithm -[Delay(New Tree) - Delay(Current Tree)]
Simulated annealing is a local neighborhood search heuristic p = exp
c
technique [25]. Two basic disadvantages of ordinary local Delay(New Tree) > Delay(Current Tree). (21)
search algorithms are that they may get stuck in local minima
because they accept only cost improving solutions and that Here, c is the control parameter that regulates the probability
the quality of the final result heavily depends on the initial of accepting a delay-increasing transition. At the beginning, a
solution. In contrast, simulated annealing algorithms occa- large value for the control parameter is chosen, resulting in the
sionally accept deteriorations in cost in a controlled manner acceptance of most of the transitions. During the search, we
besides accepting improvements in cost. This property enables slowly reduce the control parameter towards zero according
them to escape from local minima while keeping the favorable to the cooling schedule. Lower control parameter values make
features of local search algorithms, i.e., simplicity and general the acceptance of cost-increasing transitions less probable. At
applicability. any point during the run, we always keep the best spanning
Fig. 4 shows the flowchart of the annealing algorithm used tree generated up to that point.
for finding minimum-delay topologies. In general, simulated It has been shown in [26] that the simulated annealing
annealing algorithms are defined by a neighborhood structure algorithm finds the global optimum. Unfortunately, this imple-
and a cooling schedule. The neighborhood structure for our mentation requires an infinite number of transitions. A finite-
problem is defined as follows: Any two spanning tree topolo- time simulated annealing algorithm for finding high-quality
gies which have all the branches except one common are solutions can be implemented with a suitable cooling schedule.
EUSOY AND PANWAR: TOPOLOGICAL DESIGN OF INTERCONNECTED LANWAN NETWORKS 1177

A cooling schedule consists of choosing an initial value c,

for the control parameter c the method for decrementing
the control parameter, a finite number of transitions at each
value of the control parameter, and the stopping criterion.
Different cooling schedules have been proposed in [19], [25],
and [26]. We experimented with different cooling schedules

I.’\/
and chose one similar to the one described in [19], because of
its simplicity and efficiency. As already mentioned, the initial
value of the control parameter c, is chosen so that almost all
’ LAN
LAN-LAN Bridge
new topologies are accepted at the beginning. The function ,,Y LAN-MAN Bridge
used for decrementing the control parameter is given by
Fig. 5. An example topology with more than one L A N N A N bridge per
ck+1 = ff ‘ Ck = 0, 1,2, ’ . ’ cluster.

0.75 5 a 5 0.99. (22)

all communication paths between stations [ 111. LAN/MAN
The control parameter is decreased after acceptance of a fixed
bridges usually have higher capacities than ordinary bridges
number of new topologies. However, since transitions are
in order to support the cluster traffic.
accepted with decreasing probability, the number of topologies
One approach to find a minimum delay overall topology
examined at each C k will increase as Ck goes to zero. In order
is solving the problem as if it were a LAN/LAN problem
to avoid extremely long iterations at small values of C k , the with many nodes and adding an additional constraint, forcing
total number of topologies examined at each c k are bounded LAN’s in different clusters to be interconnected through the
by a fixed maximum value. This value is comparable to the
MAN. Increasing the number of LAN’s increases the running
size of the neighborhood. Simulated annealing is terminated if
time of the simulated annealing algorithm. We will instead
the value of the delay does not change after decrementing use the approach of decomposing the problem into smaller
the control parameter a fixed number of times. This fixed
problems for each cluster and, hence, reduce the running time.
number is chosen such that the algorithm has a sufficiently
In the first phase, we assume that only one LAN in each cluster
large probability of visiting at least a major part of the
is directly connected to the MAN. In other words, there is
neighborhood of a given solution. In order to guarantee that we only one LAN/MAN bridge per cluster. This is a reasonable
are not missing any good solutions in the neighborhood of the assumption because LAN/MAN bridges are currently five
annealing solution, we terminate the algorithm by comparing times more expensive than ordinary bridges. Furthermore, in
the annealing solution with all of its neighbor topologies. the case of using more than one LAN/MAN bridge, some of
the intracluster traffic has to pass through the MAN as shown
111. DESIGNOF INTERCONNECTED
LAN/MAN NETWORKS in Fig. 5 because of the spanning tree requirement. This may
In this section, we will extend the simulated annealing not be desirable for security reasons.
algorithm to the design of minimum-delay LAN/MAN net- The problem is suitable for decomposition because the
works. There are many issues to be resolved in the area of intracluster traffic does not affect other clusters. Therefore,
LAN interconnection over MAN’S, such as packet delimiting the topology of each cluster can be determined by consid-
and encapsulation, different protocols, and out-of-sequence ering only the intracluster traffic and the traffic between the
packets due to multiple routes [ 111. We will not consider these MAN and that cluster. At the beginning, we are given an
problems here, since we are only interested in the overall overall traffic requirement matrix, [ t s d ]N N , for the whole
minimum delay topology of the LAN’s interconnected over network. We can calculate smaller traffic requirement matrices
a MAN. Fig. 1 shows an example interconnected LAN/MAN [ t $ d ] ( ~ k + l ) x ( ~ kwhere
+ l ) , N k is the number of LAN’s in
network with several clusters. We know the traffic require- cluster IC. For example, if the original traffic requirements for
ments between all LAN pairs and the clustering information the topology in Fig. 1 are given by
(i.e., which LAN’s are in each cluster). We want to find a
minimum-delay LAN/MAN topology. The overall topology
has to be a spanning tree because of the transparent bridges. tl,l . . t 1 , 6
‘
We assume that we have only one MAN. This MAN can .. ..
be FDDI, IEEE 802.6 DQDB, or instead of the MAN we . .
can have a connectionless packet switched data service (e.g., t6,l
SMDS) as a backbone network. We assume that the backbone
MAN has a large capacity and the intercluster traffic will not
have significant impact on the performance of the MAN. The
backbone is modeled as a central node. The bridges between
LAN clusters and the MAN are specialized remote bridges. An tl2,l ..’
essential characteristic of remote bridging is that no station ... ...
may connect directly to the interconnecting medium, in this
case the MAN. At least two remote bridges must appear in t18,l
1178 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 11, NO. 8, OCTOBER 1993

we can calculate the traffic matrix [ti,d]for cluster 1 as follows: minimum-delay topology for that cluster as in the first phase
but, this time, the maximum number of LAN/MAN bridges is
constrained to two. If TkAdrops below the threshold, we stop.
Otherwise, we allow the cluster to have one more LAN/MAN
bridge. We continue increasing the number of LAN/MAN
bridges until TkA drops below the threshold. We repeat this
procedure for all clusters. As a result, some of the clusters
might have one LANMAN bridge, and some of them might
have more depending on their traffic requirements.
Another approach could be using additional cost constraints
to determine the number of LAN/MAN bridges, but this
would increase the complexity of the problem and currently
1=7 high LAN/MAN bridge prices (approximately five times more
After finding the new traffic requirement matrices for each expensive than LANKAN bridges) would force us to use as
cluster, we can solve each subproblem independently. At the few bridges as possible per cluster. In many practical-size
beginning, we allow only one LAN/MAN bridge per cluster. problems, one LAN/MAN bridge per cluster is sufficient and
We use the simulated annealing algorithm as in the LANKAN the overall minimum-delay topology is found in one phase.
problem with the additional constraint that each cluster will A variation of the problem occurs in the case of using SMDS
be connected to the MAN node with only one bridge. After for interconnecting the clusters of LAN’s, since SMDS will
finding the overall topology, we calculate the maximum cluster provide communications at DS1 (1.5 Mb/s) or DS3 (45 Mb/s)
access delay TkAfor each cluster. T i Ais the maximum delay rates. During the design process, we have to decide which rate
between any LAN in cluster k and the MAN in either direction. of service and, in the case of DS3, which access class (4, 10,
The maximum end-to-end delay in the overall network will 16,34 Mb/s) will be used for each cluster. A possible approach
always be bounded by the summation of the two largest T i A can be described as follows: The total traffic between each
and the delay due to the MAN ( T M A NWe ) . assumed that the cluster and the MAN is known. At first, we choose the lowest
intercluster traffic will not have a significant impact on TMAN and cheapest rate among all SMDS access classes which is
because of the large capacity of the MAN. For example, in the sufficient to carry the cluster-MAN traffic for each cluster. We
case of SMDS as a backbone MAN, it is not very likely that find the minimum-delay cluster topologies and calculate T i A .
one group of users will change the performance of the service If it is acceptable, we stop; otherwise, if TkA is large due to
because the network management system will try to maintain a the delay on LAN/SMDS bridge, we choose the next higher
guaranteed level of performance. TMAN is, therefore, assumed SMDS access class for clusters with unacceptable T i A .We
to have a fixed value for our problem. The LAN access delay can decide whether the SMDS connection is the bottleneck
between LAN i and the MAN ( T i A )is equal to the larger or not by using a threshold on the utilization of the SMDS
of the total delay in either direction between LAN i and the line. We continue increasing the SMDS rates for each cluster
MAN ( T ~ , M A N T M A N ,The
and ~ ) . delay from LAN i to the until we are satisfied with the delay due to the LAN/SMDS
MAN is equal to the summation of the delays of all LAN’s bridge and T i A .If the bottleneck is not the SMDS connection,
( p ) and bridges ( T , u ) lying on the path from LAN i to the increasing the SMDS access rates does not reduce T i A .In that
MAN, Path(i, MAN), and is given case, we can increase the number of LAN/SMDS bridges per
cluster as explained before.
Ti,MAN = E[Tp] The algorithm for finding an overall LAN/MAN topology
pEPath(i,MAN) can be summarized as follows.
0 Find individual traffic matrices for each cluster from the

given overall traffic matrix.

(r,u)EPath(i,MAN)
0 The number of LAN/MAN bridges is initially equal to

and T M A N ~ to the summation of the delays in the other one for all clusters.
is,equal
direction. Therefore, the LAN access delay and the cluster 0 For all clusters,

access delay are given by 1. Find individual minimum delay topology for the cluster.
2. Find Cluster Access Delay.
T ~ = A max ( T ~ , M A N , T M A N , ~ ) Vz, (25) 3. If Cluster Access Delay exceeds its threshold, add one
~i~ = max
, ~i~
zEcluster k
Vk (26) more LAN/MAN bridge to the cluster and go to 1;
otherwise save cluster topology and repeat 1, 2, 3 for
If TkA exceeds a threshold for a cluster, we proceed to the the next cluster.
second phase: We allow that cluster to have two LAN/MAN
bridges. Because of the spanning tree requirement, that cluster I v . RESULTSAND DISCUSSION
will have two subtrees as shown in Fig. 5. Since this is likely
to reduce the depth of the tree, the new TkA will be lower. A . Methodology for the Experiments
As we have explained before, we can change the topology of We performed experiments on networks with 6, 7, 10, 15,
one cluster without affecting the other clusters. We find the 20, and 30 LAN’s. For 6 and 7 LAN problems, we enumerated
ERSOY AND PANWAR: TOPOLOGICAL DESIGN OF INTERCONNECTED LANMAN NETWORKS 1179

TABLE I

Number of
LAN’s Tmin
Annealing
I %ax
RESULTSFOR THE MINIMUM

T L
Bound
~
Tmm
~
Local Search
~
I ~
Tmax
LAN PROBLEM
DELAYINTERCONNECTED

I
io'ooo
Tmin
smples
Tmean
I CPU Time PC-ATICONVEX
6 6.406* 6.406* 5.419 6.406* 7.214 - - 8 SI< 1 s
7 7.524* 7.746 5.721 7.524* 9.549 - - 18 SI< 1 s
10 8.323 8.524 6.199 8.639 11.027 9.103 1 1.445 52 s/< 1 s
15 10.470 10.696 7.168 11.117 14.836 12.419 15.993 6.2 min14.3 s
20 13.910 14.370 8.908 14.832 19.235 16.655 21.937 25.8 mid18 s
30 17.233 18.112 10.662 19.374 24.583 22.166 26.263 181 mi4124 s
* Global minimum in complete enumeration.

all possible spanning tree topologies and found the global value was measured on interconnected LAN’s by Leland and
minimum. This enabled us to compare the solution of the Wilson [21]. The average number of packets in a batch, X, is
simulated annealing directly with the global minimum. For 8. The capacity of LAN’s is 10 Mb/s, which is the standard
larger problems, we used several different goodness measures. for CSMA/CD LAN’s. The capacity of LANLAN bridges is
In order to estimate the range of the simulated annealing 6,000 packets/second and the capacity of LAN/MAN bridges
results, we ran the algorithm with 10 different random seeds are 10,000 packets/second for 192-byte packets. Given these
on the same problem. The difference between the best and the parameters, all the delay values presented in the following
worst solution in 10 runs gave us the range for the annealing sections are in milliseconds. The.threshold for T6A in the
solutions. We compared the solutions with the lower bound LAN/MAN problem is 20 ms.
described in Section 11-C. The gap between the lower bound
and the annealing solutions is an upper bound for the deviation B . Results for the Interconnected LAN Problem
from the global minimum. We also used the statistical good-
Table I summarizes the results of the annealing algorithm
ness measure described in [7]. In this measure, 10,000 random
and the comparisons with the lower bound for the inter-
feasible topologies were generated. A histogram corresponding
connected LAN problem in the case of heavy-load random
to the delay values of these topologies was created. This was
traffic requirements. As an example for the heavy load, the
then compared with the simulated annealing solutions.
requirement matrix for the 15 LAN problem consists of
The simulated annealing algorithm is a local search heuris-
uniformly distributed random requirements with an average
tic. In order to see its advantages over ordinary local search,
of 3 batches/second. Two rows, corresponding to LAN’s
we implemented a conventional local search heuristic and
with file servers, have higher requirements with an average
ran it several times with different random initial topologies.
10 batches/second. Corresponding minimum delay topologies
We compared the best topology found by the local search
found by the algorithm are such that maximum utilization
algorithm with the simulated annealing results. In order to on LAN’s and bridges reach 80% and 40%, respectively.
have a fair comparison, we adjusted the number of times the The minimum and maximum delay values for the simulated
local search was run so that the combined running time of annealing algorithm are found by running the algorithm with
multiple runs of local search was approximately equal to that 10 different random initial topologies. Related columns for the
of the annealing algorithm. best and worst performances of the algorithm show that the
The test problems had various traffic requirement patterns range of simulated annealing solutions is small. This small
as described. range of results confirms that the quality of the final solution
0 Three of the traffic matrices consist of uniformly dis-
is not dependent on the initial topology. For small problems,
tributed random traffic requirements with different average the algorithm found the global minimum most of the time.
values. These average values correspond to light, medium, and It found slightly higher delay topologies the rest of the time.
heavy loads. In order to consider the unbalanced inter-LAN There is a gap between the values of the simulated annealing
traffic patterns, some rows and columns of the traffic matrices solutions and the lower bound. For different size problems,
have larger average values than others. These rows or columns this gap was in the range of 18.2-61.6% of the lower bound.
might correspond to LAN’s connected to file servers or host The gap is a function of the quality of the solutions and the
computers. tightness of the lower bound. The lower bound is not very
0 One of the traffic matrices is such that the traffic between tight. For example, it can be observed from Table I that the
any two LAN’s decreases linearly with the “distance” between gap values between the lower bound and the global minimum
them. For example, the traffic requirements between LAN 1 for 6 and 7 LAN problems are 18.2% and 31.5% of the
and LAN 2 are larger than those between LAN 1 and LAN lower bound, respectively. Therefore, we conjecture that the
5. All traffic requirements have deterministic values according annealing solutions for larger problems are not very far from
to the “distance” measure. the global minimum.
0 The last traffic matrix is uniform. All traffic requirements In fact, Table I shows the worst overall performance of the
have the same deterministic value. annealing algorithm because the gap between the lower bound
Other important parameters of the example problems are as and the annealing solutions becomes smaller for the relatively
follows: The mean packet length, I, is equal to 192 bytes. This lighter load cases. The gap values were between 4.6% and
1180 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 1 1 , NO. 8, OCTOBER 1993

22.5% of the lower bound for light load cases. A complete TABLE I1
table of results for the lighter load cases is given in [27]. One
of the reasons for the gap between the solutions and the lower
bound being large is using the minimum requirement spanning
Number of Tmin ;:::1
FOR Two LAN/MAN PROBLEMS
RESULTS
ITmax
Annealing Annealing Access
CPU Time
PC-
ATICONVEX
tree [23] for finding the bound on the total flow on LANs and Cluster 1 6 8.421* 8.480 14.643 22 SI< 1 s
bridges. This tree is found without considering the capacity Cluster 2 5 7.156* 7.352 14.417 9 SI< 1 s
constraints for LAN’s and bridges and is not feasible most of Cluster 3 7 8.572* 8.674 15.553 36 SI< 1 s
the time. Thus, actual flow on a feasible tree is larger than Overall 18 9.483* 9.667 - 67 s/l s
the lower bound for the total flow. The difference between the
Cluster 1 6 8.375* 8.437 13.948 23 SI< 1 s
actual total flow and the lower bound is smaller for light load Cluster 2 12 10.929 11.258 17.663 229 sI3.6 s
cases, leading to a smaller gap. Cluster 3 7 8.430* 8.768 14.177 34 SI< 1 s
The simulated annealing algorithm outperformed multiple Cluster 4 10 9.555 9.780 16.402 86 ~11.3s
runs of the greedy local search algorithm in all cases stud- Overall 35 13.069 13.349 - 372 sl6 s
ied. The contrast between the range of simulated annealing
* Global minimum in complete enumeration
solutions and that of the greedy local search shows that
the simulated annealing algorithm does not get stuck in a
local minimum. Related columns of Table I show that the V. CONCLUSION
simulated annealing algorithm finds better topologies than the We have described a method based on simulated annealing
best of 10,OOO randomly generated topologies. Furthermore, for finding minimum-delay spanning tree topologies for in-
the majority of the random topologies have significantly higher terconnected LAN/MAN networks. We have decomposed the
delays than the simulated annealing solutions. The last column LAN/MAN topology design problem into smaller, independent
in Table I shows the average CPU times for the simulated problems for each cluster. We have derived a lower bound for
annealing algorithm on a PC/AT and CONVEX 120 minicom- the minimum-delay spanning tree problem using the solution
puter. For smaller problems, the algorithm finds high-quality of the optimum requirement spanning tree problem described
solutions very quickly even on a personal computer. For larger in [23]. Results of the computational experiments with small
problems, running times are still reasonable, given that this is problems have shown that the simulated annealing algorithm
an off-line design problem. finds the optimum topology much faster than complete enu-
meration. For larger problems, comparisons with the lower
bound have indicated that the simulated annealing solutions are
C . Results for the Interconnected LA” Problem not very far from the global optimum. Furthermore, compar-
We have experimented with two LAN/MAN interconnection isons with multiple runs of greedy local search have shown that
problems. The first problem had a small number of LAN’s the quality of the solutions found by the simulated annealing
in each cluster as shown in Fig. 1, so that the global mini- algorithm depends at most weakly on the initial topology and
mum could be found with complete enumeration. The second are better than those of ordinary local search algorithms.
problem had more clusters and more LAN’s in each cluster. Increasing demand for the interconnection of LAN’s and
The top portion of Table I1 summarizes the results for the first better LAN/MAN topologies is continuing. New backbone
LAN/MAN problem. The simulated annealing algorithm found network architectures are proposed, and the local exchange
the global optimum topology for each of the three clusters in carriers are offering connectionless data transfer services for
this problem. Since the overall topology is a combination of interconnecting LAN’s. Although they are more complex and
these cluster topologies, the solution is also globally optimal slower than bridges, routers are becoming popular. Routers
for the first problem. The short overall running time for the first are more flexible and do not require the topology to be a
problem shows the advantage of decomposing the problem. spanning tree. Design of interconnected LAN/MAN topologies
If we had approached the problem as an 18 LAN problem using routers is another problem which can be approached in
without decomposing it, the running time would have been a similar way. Since the topology is no longer required to be a
around 20 minutes instead of 67 seconds on a PC/AT. Only spanning tree, different objective functions such as minimizing
one LANMAN bridge was sufficient for each cluster in order the dollar cost and maximizing the throughput can be used.
to keep the cluster access delays below the threshold (i.e., 20 Other constraints can easily be incorporated into the formu-
ms). lation of the problem, such as constraints on the maximum
The lower portion of Table I1 summarizes the results for number of hops between any two LAN’s, maximum utilization
the second problem, which has more and larger clusters. for LAN’s, or reliability requirements. Because of its simplic-
Although we do not know the overall global optimum for ity and general applicability, simulated annealing can be used
this problem, we can still check for the global optimum for for finding good feasible solutions for many variations of the
small clusters. Goodness measures described before show that interconnected LAN/MAN topology design problem.
we have good topologies for larger clusters. Therefore, we
conjecture that the simulated annealing algorithm finds high- APPENDIXA
OPTIMUM DISTRIBUTION OF THE TOTALFLOWON LAN’s
quality solutions for the LAN/MAN problem. Running times
for the algorithm are short because of the decomposition of A lower bound for the total flow on all LAN’s (XL,,,,~~)
the problem. and a set of inequality constraints due to mandatory flows on
ERSOY AND PANWAR: TOPOLOGICAL DESIGN OF INTERCONNECTED LANNAN NETWORKS 1181

each LAN were presented in Section 11-C. The delay due to APPENDIX B
LAN’s is given by OF’TIMUM DISTRIBUTION
OF THE TOTALFLOW ON BFWGES
A lower bound for the total flow on bridges was
presented in Section 11-C. A lower bound for TB can be
obtained by optimally distributing on all bridge ports
and such that TB is minimized. This is a standard optimization
problem, which can be solved using a Lagrange multiplier.
The delay due to bridges is given by

i=l

A lower bound for TL can be obtained by finding the optimum

distribution of on all LAN’s subject to the inequality
constraints in (16) so that TL is minimized. This is a standard
optimization problem which can be solved using Lagrangean and
techniques and the following Kuhn-Tucker conditions [24]:
X;,min = Xij. (37)
VPort,]

In order to minimize TB subject to (37), we form the La-

N \
grangean L
- cpi(Xy - Xi)) =0 Vi (29)
i=l
- Xi) = 0 Vi (30)
pi 2 0 vi (31)

where a and pi’s are Lagrangean coefficients. The optimum

distribution for the general case can be found by solving the
set of equations specified by the Kuhn-Tucker conditions. where a is the Lagrange multiplier. Stationary points of C are
If all LAN’s have equal capacity (i.e., pi = p ~ , V i ) it, can found by checking where the first derivatives with respect to
be verified using the Kuhn-Tucker conditions that the water- individual Xij’s are equal to zero:
filling type distribution of described in Section 11-C and
shown in Fig. 3 minimizes TL.In this distribution, if X i # Ay,
then the corresponding pi = 0 due to (30). Consequently, (29)
becomes pij’s are known; therefore, we can find all X i j ’ s in terms of
a using (39). We also know that the summation of all X:js
is equal to from which we can find a and Xij’s. In
order to verify that these Xij’s result in a minimum for TB,
we check if the second partial derivative of C is positive
Therefore,
d2L
= 2X(Pij - X X i j ) VPOTtij. (40)
X i = Xl, vi s.t.X; 2 Ay. (33)

For LAN’s with X i = Ay, (29) becomes For normal operation, XXij < pij for all bridge ports;
otherwise, the delay will be infinite. This makes the quantity
in (40) positive, thus ensuring a minimum. Therefore, X i j ’ s
XPL
( P L - XXZ)2
-Q + pi = 0 vi S . t . X i = Ay. (34) found using this distribution give a lower bound on TL.
If all bridges have equal capacity (i.e., pij = pg,VPortij),
Therefore. the uniform distribution of among all bridge ports gives
a lower bound for TB. For this case, the following bridge port
flows give a lower bound for TB:

(35) This lower bound for TB can be improved by considering

additional inequality constraints. We pick the 2(N - 1) largest
since A; > X i . As a result, all p i ’ s satisfy (31) which shows inter-LAN traffic requirements. Since each of these require-
that we can find pi’s satisfying the Kuhn-Tucker conditions. ments has to be carried by at least one bridge, they impose
Hence, the distribution shown in Fig. 3 minimizes TL. inequality constraints on bridge flows. In order to minimize

.. .
 1182 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 11, NO. 8, OCTOBER 1993

TB, each of these requirements is assigned on a separate bridge [I91 E. Aarts and J. Korst, Simulated Annealing and Boltzmann Machines.
port. As in the case of LAN’s, finding a lower bound on TB is New York: Wiley, 1989.
[20] R. Gusella, “A measurement study of diskless workstation traffic on an
equivalent to finding the optimum distribution of on all Ethernet,” IEEE Trans. Comrnun., pp. 1557-1568, Sept. 1990.
bridge ports subject to 2(N - 1) inequality constraints due to [21] W. E. Leland and D. V. Wilson, “High time-resolution measurement
the 2 ( N - 1) largest requirements. Again, a water-filling type and analysis of LAN traffic: Implications for LAN interconnection,” in
Proc. IEEE INFOCOM, 1991, pp. 1360-1366.
distribution similar to the one shown in Fig. 3 is optimum and [22] D. Gross and C. M. Harris, Fundementals of Queueing Theory, 2nd Ed.
gives an improved lower bound on TB. New York: Wiley, 1985.
[23] T. C. Hu, “Optimum communication spanning trees,” SIAM J. Comp.,
If the bridge port capacities are not equal, the 2 ( N - 1) vol. 3, pp. 188-195, 1974.
largest requirements cannot be used for improving the lower [24] F. S. Hillier and G. J. Lieberman, Introduction to Operations Research,
bound on TB since they cannot be assigned on separate 4th Ed. New York: Holden-Day, 1986.
[25] S. Kirkpatrick, C. D. Gelatt, and M. P. Vecchi, “Optimization by
bridge ports. In that case, the lower bound found using the simulated annealing,” Sei., pp. 671-680, May 1983.
Lagrangean in (38) is used for TB.As shown in ( 5 ) , the sum [26] B. Hajek, “Cooling schedules for optimal annealing,” Math. Operat.
of the lower bounds on TL and TB gives the lower bound for Res., vol. 13, pp. 311-329, 1988.
[27] C. Ersoy, “Topological design of interconnected local and metropolitan
the average network delay T. area networks,” Ph.D. dissertation, Polytech. Univ., Brooklyn, NY, June
1992.
REFERENCES

[l] “IEEE Standard 802.1, part D: MAC Bridges,” in IEEE Project 802
Local and Metropolitan Area Network Standards. Cem Ersoy (S’87-M’91-S’91-M’92) was born in
[2] F. Backes, “Transparent bridges for interconnection of IEEE 802 Ankara, Turkey, on September 28, 1963. He re-
LAN’s,’’ IEEE Netw., pp. 5-9, Jan. 1988. ceived the B.S. and M.S. degrees in electrical engi-
[3] C. K. Kwok and B. Mukherjee, “On transparent bridging of CSMA/CD neering from Bogazici University, Istanbul, Turkey,
networks,” in Proc. IEEE GLOBECOM, 1989, pp. 185-190. in 1984 and 1986, respectively, and the Ph.D. de-
[4] L. Merakos and H. Xie, “Interconnection of CSMA/CD LAN’s via an gree in electrical engineering from the Polytechnic
N-port bridge,” in Proc. IEEE INFOCOM, 1989, pp. 2&37. University, Brooklyn, NY, in 1992.
[51 C. Ersoy, S. S. Panwar, R. Dalias, and D. Segal, “Transient phenomena
From 1984 to 1986, he worked as a Research
in bridged local area networks,” in Proc. IEEE GLOBECOM, 1990.
[6] S. C. Liang and J. R. Yee, “A gateway allocation algorithm for and Development Engineer at Northern Electric
interconnecting existing data networks,” in Proc. IEEE INFOCOM, Telecommunications (NETAS), Istanbul, Turkey. At
1989, pp. 468-473. the Polytechnic University, he worked as a Teaching
[7] P. C. Fetterolf and G. Anandalingam, “Optimal design of LAN-WAN Assistant from 1987 to 1989 and as a Research Assistant from 1989 to
internetworks: An approach using simulated annealing,” in Proc. ORSA 1992. He is currently an Assistant Professor in the Computer Engineering
Telecommun. Conf., 1990. Department of Bogazici University, Istanbul, Turkey. His research interests
[8] S. Gupta and K. W. Ross, “Performance modelling and optimization include computer networks and optimization.
of interconnected ethemets,” in Proc. IEEE INFOCOM, 1991, pp.
1353-1359.
[9] F. E. Ross, “An overview of FDDI: The fiber distributed data interface,”
IEEE J . Select. Areas Commun., pp. 1043-1051, Sept. 1989. Shivendra S. Panwar (S’82-M’85) was born in
[lo] J. F. Mollenauer, “Standards for metropolitan area networks,” IEEE
Delhi, India, on December 15, 1959. He received
Commun. Mag., pp. 15-19, Apr. 1988.
[ l 11 G. H. Clapp, “LAN interconnection across SMDS,” IEEE Netw. Mag., the B.Tech. degree in electrical engineering from
pp. 25-32, Sept. 1991. the Indian Institute of Technology, Kanpur, in 1981,
[I21 L. Mongiovi, M. Farrell, and V. Trecordi, “A proposal for interconnect- and the M.S. and Ph.D. degrees in electrical and
ing FDDI networks through B-ISDN,” in Proc. IEEE INFOCOM, pp. computer engineering from the University of Mass-
116&1167, 1991. achusetts, Amherst, in 1983 and 1986, respectively.
[13] S. E. Minzer, “Broadband ISDN and asynchronous transfer mode From 1981 to 1985, he was a Research Assistant
(ATM),” IEEE Commun. Mag., pp. 17-24, Sept. 1989. at the University of Massachusetts. He joined the
[14] P. Marsden, “Interworking IEEE 802/FDDI LAN’s via the ISDN frame Department of Electrical Engineering at the Poly-
relay bearer service,” Proc. IEEE, pp. 223-229, Feb. 1991. technic University, Brooklyn, where he is currently
[15] R. Perlman, “An algorithm for distributed computation of a spanning an Associate Professor. He spent the summer of 1987 as a Visiting Scientist at
tree in an extended LAN,” Comput. Commun. Rev., pp. 44-53, Sept. the IBM T. J. Watson Research Center, Yorktown Heights, NY, and has been
1985. a Consultant to AT&T Bell Laboratories, Holmdel, NJ. His research interests
[I61 D. Bertsekas and R. Gallager, Data Networks, 2nd Ed. Englewood include the performance analysis and design of high-speed networks. He is
Cliffs, NJ: Prentice-Hall, 1992. currently the Secretary of the Technical Affairs Council of the IEEE Commu-
[17] M. R. Carey and D. S. Johnson, Computers andIntractability: A Guide nications Society and Co-chairman of the Technical Program Committee of
to the Theory of NP-Completeness. New York: Freeman, 1979. the Second IEEE Network Management and Control Workshop.
[18] W. Nijenhuis, Combinatorial Algorithms. New York Academic, 1975. Dr. Panwar is a member of Tau Beta Pi and Sigma Xi.

View publication stats