Impact of Interference on Multi-hop Wireless Network

Performance

Kamal Jain Jitendra Padhye Venkat Padmanabhan Lili Qiu

Microsoft Research
One Microsoft Way, Redmond, WA 98052.
fkamalj, padhye, padmanab, [email protected]

ABSTRACT 1. INTRODUCTION
In this paper, we address the following question: given a Multi-hop wireless networks have been studied since the
specic placement of wireless nodes in physical space and a 70's [6]. Several new applications of such networks have
specic traÆc workload, what is the maximum throughput recently emerged. Community wireless networks [1, 3] are
that can be supported by the resultant network? Unlike multi-hop wireless networks that provide \last-mile" access
previous work that has focused on computing asymptotic to peoples' homes. This approach is an alternative to to
performance bounds under assumptions of homogeneity or cable modem and DSL technologies. In large networks of
randomness in the the network topology and/or workload, sensors [10] the scale and the environment are such that
we work with any given network and workload specied as a multi-hop wireless network is the only feasible means of
inputs. communication.
A key issue impacting performance is wireless interference A fundamental issue in multi-hop wireless networks is that
between neighboring nodes. We model such interference us- performance degrades sharply as the number of hops tra-
ing a con ict graph, and present methods for computing versed increases. For example, in a network of nodes with
upper and lower bounds on the optimal throughput for the identical and omnidirectional radio ranges, going from a sin-
given network and workload. To compute these bounds, we gle hop to 2 hops halves the throughput of a ow because
assume that packet transmissions at the individual nodes wireless interference dictates that only one of the 2 hops can
can be nely controlled and carefully scheduled by an omni- be active at a time.
scient and omnipotent central entity, which is clearly unreal- The performance challenges of multi-hop networks have
istic. Nevertheless, using ns-2 simulations, we show that the long been recognized and have led to a lot of research on
routes derived from our analysis often yield noticeably bet- the medium access control (MAC), routing, and transport
ter throughput than the default shortest path routes even layers of the networking stack. In recent years, there has
in the presence of uncoordinated packet transmissions and also been a focus on the fundamental question of what the
MAC contention. This suggests that there is opportunity for optimal capacity of a multi-hop wireless network is. The
achieving throughput gains by employing an interference- seminal paper by Gupta and Kumar [16] showed that in a
aware routing protocol. network comprising of n identical nodes, each of which is
communicating with1 another node, the throughput capacity
Categories and Subject Descriptors per node is ( p log ) assuming random node placement
n n

C.2.1 [Computer-Communication Networks]: Wireless and communication pattern and ( p1 ) assuming optimal
networks|multi-hop, interference node placement and communication pattern. Subsequent
n

work has considered alternative models and settings, such

General Terms as the presence of relay nodes and mobile nodes, and local-
Algorithm, Performance ity in inter-node communication, and their results are less
pessimistic [13, 20, 12].
This paper also deals with the problem of computing the
Keywords optimal throughput of a wireless network. However, a key
wireless, multi-hop, interference, performance distinction of our work from previous work such as [16] is
that we work with any given wireless network conguration
and workload specied as inputs. In other words, the node
locations, ranges, etc. as well as the traÆc matrix indicat-
Permission to make digital or hard copies of all or part of this work for ing which source nodes are communicating with which sink
personal or classroom use is granted without fee provided that copies are nodes are specied as the input. We make no assumptions
not made or distributed for profit or commercial advantage and that copies about the homogeneity of nodes with regard to radio range
bear this notice and the full citation on the first page. To copy otherwise, to
or other characteristics, or regularity in communication pat-
republish, to post on servers or to redistribute to lists, requires prior specific
tern. This is in contrast to previous work that has focused
on asymptotic bounds under assumptions such as node ho-
permission and/or a fee.
MobiCom’03, September 14–19, 2003, San Diego, California, USA.
Copyright 2003 ACM 1-58113-753-2/03/0009 ... 5.00. $
mogeneity and random communication patterns. ours.
We use a con ict graph to model the eects of wireless In their seminal paper [16], Gupta and Kumar studied the
interference. The con ict graph indicates which groups of capacity of wireless networks under two models of interfer-
links mutually interfere and hence cannot be active simulta- ence: a protocol model that assumes interference to be an all-
neously. We formulate a multi-commodity ow problem [8], or-nothing phenomenon and a physical model that considers
augmented with constraints derived from the con ict graph, the impact of interfering transmissions on the signal-to-noise
to compute the optimal throughput that the wireless net- ratio. They show that in a network comprising of n identical
work can support between the sources and the sinks. We nodes, each of which is communicating with1 another node,
show that the problem of nding optimal throughput is NP- the throughput capacity per node is ( p log ) assuming
hard, and we present methods for computing upper and n n

random node placement and ( p1 ) assuming optimal node

lower bounds on the optimal throughput.
placement and communication pattern. These results are
n

We show how our methodology can accommodate a diver-

sity of wireless network characteristics such as the availabil- shown under the protocol model, but the latter result also
ity of multiple non-overlapping channels, multiple radios per holds in the case of the physical model under reasonable as-
node, and directional antennae. We also show how multiple sumptions. According to the intuitive explanation in [20],
MAC protocol models as well as single-path and multi-path while the overall one-hop capacity of the network
p grows as
routing constraints can be accommodated. O(n), the average path length grows as O( n), so the end-
We view the generality of our methodology and the con- to-end throughput per node is O( p1 ).
Li et al. [20] have extended the work of Gupta and Ku-
n

ict graph framework as a key contribution of our work.

To compute bounds on the optimal throughput, we as- mar [16] by considering the impact of dierent traÆc pat-
sume that packet transmissions at the individual nodes can terns on the scalability of per node capacity. They point
be nely controlled and carefully scheduled by an omniscient out that a random traÆc pattern represents the worst case
and omnipotent central entity. While this is clearly an un- from the viewpoint of per-node capacity. They also show
realistic assumption, it gives us a best case bound against that for traÆc patterns with power law distance distribu-
which to compare practical algorithms for routing, medium tions, the per-node capacity stays roughly constant as the
access control, and packet scheduling. Moreover, ns-2 simu- network size grows, provided the distance distribution de-
lations show that the routes derived from our analysis often cays more rapidly than the square of the distance. Li et
yield noticeably better throughput than the default short- al. also consider the interactions of packet forwarding with
est path routes, even in the presence of real-world eects the 802.11 MAC and show that the use of 802.11 instead of
such as uncoordinated packet transmissions and MAC con- a global scheduling scheme does not aect the asymptotic
tention. In some cases, the throughput gain is over a factor bound on per-node capacity derived in [16].
of 2. The reason for this improvement is that in optimizing In [13], Grossglauser and Tse introduce mobility into the
throughput, we tend to nd routes that are less prone to model presented in [16], and show that the average long-term
wireless interference. For instance, a longer route along the throughput per source-destination pair can be kept constant
periphery of the network may be picked instead of a shorter even as the number of nodes per unit area increases, pro-
but more interference prone route through the middle of the vided we allow for delays of the order of the time-scale of
network. mobility. This is achieved by exploiting mobility to keep
We use our technique to evaluate how the per-node through- data transfers local, and transmitting only when the trans-
put in a multi-hope wireless network varies as the number mitter and receiver are close to each other, at a distance of
of nodes grows. Previous work (e.g., [16]) suggests that the O( p1 ), thereby reducing total resource usage and interfer-
per-node throughput falls as the number of nodes grows. ence. While this is encouraging, in many practical situations
n

But this result is under the assumption that nodes always (such as community wireless networks), mobility may be too
have data to send and are ready to transmit as fast as their infrequent (or even non-existent) to be exploitable.
wireless connection will allow. In a realistic setting, however, Gastpar and Vetterli [12] extend the work of Gupta and
sources tend to be bursty, so nodes will on average transmit Kumar [16] in a dierent direction. Instead of the simple
at a slower rate than the speed of their wireless link. In such point-to-point coding assumption made in [16] (which basi-
a setting, we nd that the addition of new nodes can actually cally treats each transmitter-receiver pair as being indepen-
improve the per-node throughput because the richer connec- dent of other pairs), they consider a network coding model
tivity provides increased opportunities for routing around where nodes could cooperate in arbitrary ways (for instance,
interference \hotspots" in the network. This more than o- to boost the transmit power). Further, they assume that
sets the increase in traÆc load caused by the new node. there is a single source and single destination picked at ran-
The rest of this paper is organized as follows. In Section 2, dom, and that the rest of the nodes act as relays. They show
we discuss related work. In Section 3, we present details of that the throughout capacity of the network under these
our con ict graph model and methods for computing bounds conditions is O(log n) (compared to O(1) for the point-to-
on the optimal network throughput. In Section 4, we present point coding model of [16]). While the use of network coding
results obtained from applying our model to dierent net- in this context is a promising line of research, we note that
work and workload congurations. Section 5 concludes the the point-to-point coding model corresponds to current ra-
paper. dio technology such as 802.11.
The recent work of De Couto et al. [9], based on two exper-
2. RELATED WORK iments in a 802.11b-based multi-hop wireless testbeds shows
A number of papers have been published on the problem that minimizing the hop count of an end-to-end path is not
of estimating the capacity of a multi-hop wireless network. suÆcient for achieving good performance. The reason they
Here, we consider the work that is most closely related to point out is that link quality can vary widely and that the
long hops that may be included in \short" paths may in- Physical Model: Suppose node n wants to transmit to
i

cur a high packet error rate. In our work, we also reach node n . We can calculate the signal strength, SS , of
j ij

the same conclusion regarding the limitations of the hop n 's transmission as received at n . The transmission is
i j

count metric, but for a somewhat dierent reason | be- successful if SNR  SNR
ij , where SNR denotes
thresh ij

cause of wireless interference limits capacity, a circuitous the signal-to-noise ratio at the node n for transmissions j

but less interference-prone route, say along the periphery of received from node n . The total noise, N , at n consists
i j j

a network, may perform better than the shortest hop count of the ambient noise, N , plus the interference due to other
a

route. ongoing transmissions in the network. Note again that there

Finally, Yang and Vaidya [25] introduce the notion of is no requirement that the noise level at the sender also be
a \con ict graph" in the context of their work on prior- low.
ity scheduling in wireless ad hoc networks. However, the Our goal is to model wireless interference using a general
con ict graph is dened on ows rather than links, and is framework that would then enable us to compute the opti-
used only to interpret experimental results showing that the mum throughput that the wireless network can support for
802.11 MAC causes ows with a high degree of con ict to a given traÆc workload. We assume that packet transmis-
suer disproportionately compared to ows with a low de- sions at the individual nodes can be nely controlled and
gree of con ict. There is no attempt to analyze the con ict carefully scheduled by an omniscient and omnipotent cen-
graph derive throughput bounds. tral entity. So here we do not concern ourselves with issues
In summary, there is a large body of work on the wireless such as MAC contention that may be unavoidable in a dis-
capacity problem, much of it focused on asymptotic bounds tributed network; such real-world issues are considered in
under assumptions such as node homogeneity and random the ns-2 simulation results presented in Section 4. In our
communication patterns. In contrast, our work focuses on problem formulation here, we focus on maximizing the total
computing throughput bounds for a given wireless network throughput between source-destination pairs assuming that
and traÆc workload, using a con ict graph to model the the sources always have data to send and the destination
constraints imposed by wireless interference. We do not nodes are always ready to accept data. Our framework car-
consider how factors such as mobility [13] or coding [12]. ries over easily to a dierent objective function such that one
And like [16], we do not compute the information theoretic tries to maximize the minimum throughput across all source-
capacity of the network. destination pairs and thereby achieves a degree of fairness.
Moreover, we can also accommodate the case where the rate
3. COMPUTING BOUNDS ON OPTIMAL at which nodes generate data or are willing to accept data
is bounded simply by creating a virtual source or sink node
THROUGHPUT and connecting it to the real source or sink via a virtual
We now present our framework for incorporating the con- link of speed equal to the source or sink rate. The virtual
straints imposed by interference in a multi-hop wireless net- link is special in that it is assumed not to interfere with any
work and then present methods for computing bounds on other link in the network. The virtual link is just convenient
the optimal throughput that a give network can support for construct to help us model the bound on the source or sink
a given traÆc workload. We begin with some background rate.
and terminology. The communication between the sources and destinations
can be either direct or be routed via intermediate nodes. We
3.1 Background and Terminology say that a network throughput D is feasible if there exists a
Consider a wireless network with N nodes arbitrarily lo- schedule of transmissions such that no two interfering links
cated in a plane. Let n ; 1  i  N denote the nodes, and d
i ij are active simultaneously, and the total throughput for the
denote the distance between nodes n and n . Each node,
i j given source-destination pairs is D.
n , is equipped with a radio with communication range R In the rest of this section, we consider the following three
and a potentially larger interference range R0 . For ease of
i i

scenarios in detail: (i) multipath routing under the protocol

exposition, we assume that there is a single wireless channel.
i

interference model, (ii) multipath routing under the phys-

We consider two models, the Protocol Model and the Physi- ical interference model, and (iii) single-path routing under
cal Model , to dene the conditions for a successful wireless both models. We end the section by discussing several other
transmission. These models are similar to those introduced generalizations, and provide a summary description of our
in [16]. framework.
Protocol Model: In the protocol model, if there is a single
wireless channel, a transmission is successful if both of the 3.2 Multipath Routing under the Protocol In-
following conditions are satised: terference Model
1. d  R
ij i Given a wireless network with N nodes, we rst derive
2. Any node n , such that d  R0 , is not transmitting a connectivity graph C as follows. The vertices of C corre-
k kj k
spond to the wireless nodes (N ) and the edges correspond
C

Note that the second requirement implies that a node may to the wireless links (L ) between the nodes. There is a di-
C

not send and receive at the same time nor transmit to more rected link l from node n to n if d  R and i 6= j . We
ij i j ij i

than one other node at the same time. Note also that this use the terms \node" and \link" in reference to the connec-
model diers from the popular 802.11 MAC in an important tivity graph while reserving the terms \vertex" and \edge"
way | it requires only the receiver to be free of interference, for the con ict graph presented in Section 3.2.1.
instead of requiring that both the sender and the receiver Let us rst consider communication between a single source,
be free of interference. We discuss how to adapt the model n , and a single destination, n . In the absence of wireless
s d

for an 802.11-style MAC in Section 3.5. interference (e.g., on a wired network), nding the maxi-
 mum achievable ow between the source and the destina- Theorem 1. Given a network and a set of source and
tion, given the exibility of using multiple paths, can be destination nodes, it is NP-hard to nd the optimal through-
formulated as a linear program corresponding to a max- ow put under the protocol interference model. Moreover, it is
problem, as shown in Figure 1. Here, f denotes the amount ij NP-hard to approximate the optimal throughput.
of ow on link l , Cap denote the capacity of link l , and
It can be shown that the problem of nding the inde-
ij ij ij

L is a set of all links in the connectivity graph.

pendence number of a graph (a known hard problem even
C

max
X fsi
to approximate) can be reduced to the optimal throughput
problem. Moreover, this reduction is approximation pre-
si 2 C serving. Hence the above hardness result. We discuss the
X to X reduction brie y in the Appendix.
l L

subject
fij = fji 8n 2 N n fn ; n g <1> Since it is NP-hard to approximate the optimal through-
put, we now look at heuristics for obtaining lower and upper
Xf
i C s d

l ij 2LC ji 2LC
l
bounds on the throughput. For this, we need to dene some
= 0 <2> more terms. An independent set of a graph H can be char-
isX
is

l 2 C
L acterized using an independence vector, which is a vector of
f = 0 <3> size jV j. This vector is denoted by x , where I is an in-
H I

dependent set. The j element of this vector is set to 1 if

di

di 2 C
th

and only if the vertex v is a member of the independent

l L

fij  Cap 8l 2 L <4> j

set I , otherwise it is zero. We can think of x as a point in

ij ij C

fij  0 8l 2 L
ij C <5>
I

an jV j-dimensional space. The polytope dened by convex

H

combination of independence vectors is called the indepen-

Figure 1: LP formulation to optimize the through- dent set polytope (also known as the stable set polytope).
put for a single source-destination pair. 3.2.3 Lower Bound
The maximization states that we wish to maximize the The problem of deriving a lower bound is equivalent to
sum of ow out of the source. The rst constraint represents the problem of nding a network throughput D that has a
ow-conservation, i.e., at every node, except the source and feasible schedule to achieve it. We make the following ob-
the destination, the amount of incoming ow is equal to servation. Links belonging to a given independent set in
the amount of outgoing ow. The second constraint states con ict graph F can be scheduled simultaneously. Suppose
that the incoming ow to the source node is 0. The third there are a total of K maximal independent sets in graph F .
constraint states that the outgoing ow from the destination (A maximal independent set is one that cannot be grown fur-
node is 0. The fourth constraint indicates the amount of ther.) Let I1 ; I2 ; : : : I denote these maximal independent
K

ow on a link cannot exceed the capacity of the link. The sets, and  ; 0    1 denote the fraction of time allo-
i i

nal constraint restricts the amount of ow on each link to cated to the independent set I (i.e., the time during which
i

be non-negative. The above formulation does not take into the links in I can be active). If we add the schedule restric-
i

account wireless interference, which may impose additional tions imposed by the independent sets to the original linear
constraints on when the individual wireless links may be program (Figure 1), the resulting throughput always has a
active. We turn to this issue next. feasible schedule, and therefore constitutes a lower bound
on the maximum achievable throughput.
3.2.1 Conflict Graph We formalize our above observation as follows. Given a
To incorporate wireless interference into our problem for- con ict graph F , we dene a usage vector, U , of size jV j,
F

mulation, we dene a con ict graph, F , whose vertices cor- where U denotes the fraction of time that the link i can be
i

respond to the links in the connectivity graph, C . There is active. A usage vector is schedulable if the corresponding
an edge between the vertices l and l in F if the links l links can be scheduled, con ict free, for the fraction of the
ij

and l may not be active simultaneously. Based on the pro-

pq ij
time indicated in the usage vector. If we think of the usage
pq

tocol interference model described in Section 3.1, we draw vector as a point in an jV j-dimensional space, we have the
F

such an0 edge if any of the following is true: d  R0 or following theorem, the proof for which is not included due
d  R . This encompasses the case where a con ict arises
pj
iq i
to space constraints.
because links l and l have a node in common (i.e., i == p
p

ij pq

or i == q or j == p or j == q). Note, however, that we Theorem 2. A usage vector is schedulable if and only
if it lies within the independent set polytope of the con ict
do not draw an edge from a vertex to itself in the con ict graph.
graph.
Theorem 2 implies that the optimal network throughput
3.2.2 Hardness Result problem is a linear program, no matter how many sender-
Next, we present a hardness result for computing the opti- receiver pairs we have. In fact, the problem is one of max-
mal throughput under the protocol interference model. We imizing a linear objective function over a feasible polytope.
begin with a few denitions, Given a graph H (with ver- This feasible polytope can be described as the intersection
tex set V ), an independent set is a set of vertices, such
H of two polytopes | the ow polytope and the independent
that there is no edge between any two of the vertices. The set polytope of the con ict graph. The ow polytope is the
independence number of graph H is the size of the largest collection of feasible points described by the ow constraints
independent set in H . Then, we have the following hardness (Figure 1), ignoring wireless con icts. The ow polytope is
result. a simple structure on which a linear objective function can
easily be optimized. Independent set polytope, on the other 1 1

hand, is a diÆcult structure and no characterization of it is

known (because there may be exponentially many indepen- 5 2 5 2

dent sets).
Theorem 2 implies that any convex combination of in-
dependence vectors is schedulable. In general, however, an
arbitrary point inside the independent set polytope will be a 3 4 3

convex combination of an exponentially many independence

4

vectors. To get around this computational problem, we only Figure 2: A pentagon and its complement
want to pick \easy" points in the independent set polytope. graph. The former is an odd hole, and the
An obvious notion of \easy" is that the point picked should latter is an odd anti-hole.
be a convex combination of a small number of (i.e., polyno-
mially many) independence vectors. We will be using this
notion explicitly in the algorithm as follows. We derive a
lower bound on the optimal throughput by nding K 0 in-
dependence vectors in the con ict graph F , and adding the
following constraints to the LP formulation shown in Fig-
ure 1.
P
 =1   1 (because only one maximal independent
K
0
i Figure 3: An example that shows it is
set can be active at a time)
i

ij
P
 f  ij 2 i  Cap (because the fraction of time for
l I i ij
not suÆcient even if we add all clique,
hole, anti-hole constraints.
which a link may be active is constrained by the sum
of the activity periods of the independent sets it is a corresponding to odd holes and odd anti-holes. An odd hole
member of). is a circle (i.e., loop) formed by an odd number of edges,
Note the solution produced by solving this linear program without a chord in between. (The pentagon in Figure 2 is
is always feasible (i.e., schedulable). This is due to the fact an odd hole.) The sum of the link utilization in an odd hole
that all links belonging to independent set I can be simulta- containing k vertices can be no more than b 2 c. An odd anti-
k

hole is the complementary graph of an odd hole. Figure 2

that the P =1   1. Moreover, theorem 2 assures us that
i

neously active for  fraction of time, and we have required

shows an example of an anti-hole with 5 nodes. The sum of
i
0
K
i

when we include all independent sets, the solution will be ex-

i
link utilizations in an odd anti-hole can be no more than 2.
act, i.e., this will be the maximum value of D that is feasible. Unfortunately, even if we consider the constraints imposed
To help tighten the lower bound more quickly, we should by the odd holes and odd anti-holes (in addition to those
consider using maximal independence sets. While nding imposed by the cliques), we are not guaranteed to have a
all maximal independent sets is also NP-complete [11], the feasible solution. For example, consider the con ict graph,
lower bound obtained by considering a subset of the (max- as shown in Figure 3. We can assign a utilization of 0.4
imal) independent sets has the nice property that as we to all the vertices on the pentagon and 0.2 to the center of
add more constraints, the bound becomes tighter, eventu- the pentagon, while satisfying all clique, hole, and anti-hole
ally converging to the optimal (i.e., the maximum feasible constraints. But there is no feasible schedule to achieve this,
bound) when we add all the constraints. because this solution does not lie in the stable-set polytope.
In fact, the upper bound based only on clique constraints is
3.2.4 Upper Bound tight only for a special class of con ict graphs called perfect
In this section, we derive an upper bound on the network graphs. Perfect graphs are the graphs without any odd holes
throughput. Consider the con ict graph. A clique in the or odd anti-holes. Thus, in our present formulation, the
con ict graph is a set of vertices that mutually con ict with upper bounds may not always be tight. We will discuss this
each other. Theorem 2 implies that the total usage of the further in Section 5.
links in a clique is at most 1. This gives us a constraint on
the usage vector. We can nd many cliques and write cor- 3.3 Multipath Routing under the Physical In-
responding constraints to dene a polytope. We can then terference Model
maximize the throughput over the intersection of this poly- As before, we begin by creating a connectivity graph C ,
tope with ow polytope. This will give us an upper bound whose vertices correspond to the nodes in the network. Based
on the throughput. on the physical interference model, there exists a link, l , ij

Unfortunately, it is computationally expensive to nd all from n to n if and only if SS =N  SNR

i j ij a (i.e., the
thresh

the cliques, and even if we could nd them all, there is still SNR exceeds the threshold at least in the presence of just
no guarantee that our upper bound will be tight. This can the ambient noise).
be illustrated by the following example. Suppose the con ict Using the connectivity graph, we can write an LP formu-
graph is the pentagon depicted in Figure 2. As we can see, lation to optimize network throughput for a wired network.
the only cliques in the graph are formed by the adjacent As discussed before, the solution to the linear program, as
pairs of nodes. Adding the clique constraints alone to the shown in Figure 1, provides an upper bound on network
LP would suggest that a sum of link utilization equal to 2.5 throughput. However, this bound is not very useful since it
is achievable. But actually at most 2 links can be active does not take interference eects into account.
at a time. This suggests that we need to add constraints To take interference eects into account, we construct a
con ict graph F . Unlike in the protocol model, con icts in U1 + U2 + :::Ut + Ua  t, where as before U denotes the i

the physical model are not binary. Rather, the interference fraction of time for which physical link l (corresponding
i

gradually increases as more neighboring nodes transmit, and to vertex v in the con ict graph) is active. By adding as
i

becomes intolerable when the noise level reaches a thresh- many such constraints as possible, we can tighten the upper
old. This gradual increase in interference suggests that we bound. Still, the bound is not guaranteed to converge to the
should have a weighted con ict graph, where the weight of optimal even if we include all such sets.
a directed edge from vertices l to vertices l (denoted by
pq ij

w ) indicates what fraction of the maximum permissible

pq
3.4 Single-path Routing
noise at node n (for link l to still be operational) is con- So far we have considered multipath routing. As many
ij
j ij

tributed by activity on link l (i.e., node n 's transmission

pq p existing routing algorithms [18, 24, 23, 22] are conned to
to node n ). Specically, we have
q single-path routing, it is useful to derive a throughput bound
SS for single-path so that we can compare how much the current
w = ij
pq pj
protocols deviate from the theoretical achievable throughput
Na under the same routing restriction. The way we enforce the
ij SS
SN R thresh
where SS and SS denote the signal strength at node single-path restriction for the ow from a source to a desti-
pj ij

n of transmissions from nodes p and i, respectively, and nation is by adding the following additional constraints to
j
ij
SS
N is the maximum permissible interference
the LP problem for the wired network (shown in Figure 1):
thresh a

noise at node n that would still allow successful reception  For each link l , f  Cap
z , where z 2 f0; 1g
SN R

P
j

of node n 's transmissions. The edges of the con ict graph

ij ij ij ij i;j

 At each node n , z  1
i

are directed, and in general w may not be equal to w . pq

ij
ij
pq i ij

3.3.1 Lower Bound Here z is a 0{1 variable that indicates whether or not
ij

In the protocol model, we derive a lower bound on the net- link l is used for transmissions, and f is the amount of
ij ij

work throughput by nding independent sets in the con ict ow on the link. The basic intuition for these constraints is
graph F , and adding the constraints associated with the in- that in a single-path routing, at any node in the network,
dependent sets to the LP for the wired network. Analogous there is at most one out-going edge that has a non-zero ow.
to independent sets, we introduce the notion of schedulable The rst and the third condition ensure that at node n at i

sets in the physical model. A schedulable set H is dened most one z will have a value of 1.
ij

In theory, solving integer linear program is a NP-hard [11],

P a 2set ofw vertices
x

as such that for every vertex l 2 H , ij x

but in practice, programs such as lp solve [7] and CPLEX [5]
pq x  1. It follows that all links in a schedulable
pq
ij
can solve mixed-integer problems.
set can be active simultaneously. Suppose we schedule the
l H

links belonging to H for time  ; 0    1. We now

x x x 3.5 Other Generalization
take the original LP for the wired network (in Figure 1), The basic con ict graph model is quite exible, and can
and include the following constraints: be generalized in many ways.
 P =1   1, where K 0 is the number of schedulable
K
0
x
Multiple Source-Destination Pairs: We can extend our
formulations in the previous sections from a single source-
sets found
x

P
 f  ij 2 x  Cap
destination pair to multiple source-destination pairs using a
multi-commodity ow formulation [8] augmented with con-
straints derived from the con ict graph. We assign a connec-
ij x ij
l H

To tighten the bound, we should consider using maximal tion identier to each source-destination pair. Instead of the
schedulable sets in graph F (i.e., a schedulable set such that ow variables f , we introduce the variable f to denote
ij ijk

adding additional vertices to the set will violate the schedu- the amount of ow for connection k on link l . Referring ij

lable property). We have the following theorem, which is to Figure 1, the ow conservation constraints at each node
similar to the Theorem 2 in the protocol model. apply on a per-connection basis (constraint <1>); the total
incoming ow into a source node is zero only for the connec-
Theorem 3. A usage vector is schedulable if and only if tion(s) originating at that node (constraint <2>); likewise,
it lies in the schedulable set polytope of the con ict graph. the total outgoing ow from a sink node is zero only for the
connection(s) terminating at that node (constraint <3>);
3.3.2 Upper Bound and the capacity constraints apply to the sum of the ow
To derive an upper bound, we consider maximal sets of levels of all connections traversing a link (constraint <4>).
vertices in F such that for any pair of vertices l and l , pq ij This generalization is applicable to all variants of the
w  1. These correspond to the cliques in the protocol
pq
problem.
interference model. Therefore for each such set, we add a Multiple wireless channels: It may be the case that in-
ij

constraint that the sum of their utilization has to be no more stead of just one channel, each node can tune to one of M
than 1. channels, M  1. This can be easily modeled by introducing
These constraints may result in a loose bound since there M links between nodes i and j , instead of just 1. In gen-
may not be very many cliques. To tighten the upper bound, eral, links corresponding to dierent channels do not con ict
we further augment the linear program with the following with each other, re ecting the fact that the channels do not
additional constraints. After we nd a maximal schedula- mutually interfere. However, the links emanating from the
ble set, say vertices v1 , v2 , ..., v , adding any additional
t same node do con ict, re ecting the constraint that the sin-
vertex, denoted as v , to the set will make the set un-
a gle radio at each node can transmit only on one channel at
schedulable. Therefore we have the following constraint: a time.
Multiple radios per node: Each wireless node may be 6 16
equipped with more than one radio. If each node has M 2 5 8
radios, this can be modeled by introducing M links between 14 22
each pairs of nodes. If we assume that each of these radios
is tuned to a separate channel, and that a node can commu- 3 5 12 15 21 23

nicate on multiple radios simultaneously, then the con ict

graph will show no con ict among the M links between a
4 13

pair of nodes. 1 4 7
Directional antennas: We can combine the use of direc- 10 19
tional antennas with the basic protocol model of communica- 11 18 20
tion. Instead of specifying a range for each node, we simply 0 2 8

specify a list of nodes (or points in space) where transmis-

sions or interference from this node can be perceived. The
1 9

connectivity graph and the con ict graph are modied to 0 3 6

take this into account. 7 17

Other models of interference: In the simple example, Figure 4: 3x3 Grid

we considered an optimistic model of interference that did
not require the sender to be free of interference. But a more
realistic model, which more closely re ects the situation in shown that as we include more distinct independent sets, the
802:11, would require both the sender and the receiver to lower bound becomes progressively tighter. In other words,
be free of interference. This re ect the fact that 802:11 may the more eort we spend looking for independent sets in our
perform virtual carrier sensing using an RTS{CTS exchange, con ict graph, the better the bound will be. Since we can
and that for successful communication, the sender must be not always hope to nd optimal solutions, any upper or lower
able to hear the link layer acknowledgment transmitted by bounds discovered by our model need to be presented along
the receiver. Therefore, we draw an edge in the con ict with the amount of eort required to nd those bounds.
graph between vertices l and l if d  R0 for ab = Thus we require a metric to measure this eort. We use the
following simple algorithm to nd distinct independent sets:
ij pq ab

iq; qi; ip; pi; jp; pj; jq; or qj .

a

3.6 Summary 1. Start with an empty independent set IS.

In this section, we presented the concept of a con ict 2. Consider a random ordering of vertices in the con ict
graph, and discussed how it could be used to derive upper graph.
and lower bounds on the optimal throughput that a wireless
network can support, for a given set of sources and destina- 3. Consider the vertices of the graph in that order. Al-
tions. We also generalize the con ict graph to handle a wide ways add the rst vertex to IS.
range of scenarios. We have shown that the lower bound de-
rived from our framework is always schedulable, and will be 4. Add a new vertex if and only if it does not have an
optimal once all the independent sets are found. If the up- edge to any of the vertices added to IS so far (i.e., if
per and lower bounds are equal, then these correspond to it does not con ict with any of the existing vertices).
the optimal solution. Once we consider all the vertices, IS will be of size at
least one.
4. RESULTS 5. We check to see if we have previously discovered this
This section presents several results based on our model. independent set, and if not, we add constraints based
The section is organized as follows. In Section 4.1, we on this independent set to our linear program. Other-
present illustrative results that demonstrate the exibility of wise we discard the set.
our model. In Section 4.2, we use our model to provide in-
sights into the tradeo between the richer connectivity pro- We consider this entire sequence as one unit of eort. Note
vided by increasing the size of a wireless mesh network, and that one unit of eort does not always result in addition of
the increase in cumulative traÆc load due to the new mesh a constraint or variable to the linear program. Moreover,
participants. In Section 4.3, we illustrate how optimal rout- there is a complex relationship between the number of vari-
ing can bring benets even in absence of optimal scheduling ables and constraints in a linear program, and the amount
(i.e., in the presence of MAC contention and other ineÆ- of time required to solve it. Thus, the metric is only a rough
ciencies). In Section 4.4, we discuss the issue of convergence guide for amount of actual time (or CPU cycles) spent while
of the upper and lower bounds on the optimal throughput. nding the bound. In Section 4.5, we will provide further
Finally, in Section 4.5, we present a discussion of the com- discussion about the relationship between the eort metric,
putational costs of our model. and actual time spent in computation. The eort metric is
dened in a similar manner by considering cliques in case of
4.1 Illustrative results searching for the upper bound, and by considering schedu-
In this section, we present several illustrative results to lable sets in case of the physical model.
demonstrate the capabilities of our model. We begin by We consider the topology shown in Figure 4. The net-
dening a metric for computational eort. In Section 3, we work consists of 9 nodes, placed in a 3x3 grid. We make no
have described the procedure for nding upper and lower claims that this topology is representative of typical wire-
bounds on throughput. Let us consider the protocol model less networks. We have deliberately chosen a small, simple
of interference, and let us focus on the lower bound. We have topology, to facilitate detailed discussion of the results.
 link 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0 1 0 1 0 0 0 0
1 1 0 1 1 1 1 0 1 1 1 1 1 1 1 0 1 0 1 1 1 1 0 0 0
2 1 1 0 1 1 1 1 1 1 1 1 1 1 1 1 1 0 1 0 1 0 0 0 0
3 1 1 1 0 1 1 1 1 1 0 1 1 1 1 1 1 1 0 0 1 0 0 1 0
4 1 1 1 1 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
5 1 1 1 1 1 0 1 1 1 0 1 1 1 1 1 1 1 0 0 1 0 0 1 0
6 1 0 1 1 1 1 0 0 1 0 1 1 1 1 1 1 1 0 0 1 0 1 1 1
7 1 1 1 1 1 1 0 0 1 1 1 1 1 1 0 1 0 1 1 1 1 0 0 0
8 1 1 1 1 1 1 1 1 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
9 1 1 1 0 1 0 0 1 1 0 1 1 1 1 0 1 0 1 1 1 1 1 0 1
10 1 1 1 1 1 1 1 1 1 1 0 1 1 1 1 1 1 1 1 1 1 1 1 1
11 1 1 1 1 1 1 1 1 1 1 1 0 1 1 1 1 1 1 1 1 1 1 1 1
12 1 1 1 1 1 1 1 1 1 1 1 1 0 1 1 1 1 1 1 1 1 1 1 1
13 1 1 1 1 1 1 1 1 1 1 1 1 1 0 1 1 1 1 1 1 1 1 1 1
14 1 0 1 1 1 1 1 0 1 0 1 1 1 1 0 1 1 0 0 1 0 1 1 1
15 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0 1 1 1 1 1 1 1 1
16 0 0 0 1 1 1 1 0 1 0 1 1 1 1 1 1 0 0 1 1 1 1 1 1
17 1 1 1 0 1 0 0 1 1 1 1 1 1 1 0 1 0 0 1 1 1 1 0 1
18 0 1 0 0 1 0 0 1 1 1 1 1 1 1 0 1 1 1 0 1 1 1 1 1
19 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0 1 1 1 1
20 0 1 0 0 1 0 0 1 1 1 1 1 1 1 0 1 1 1 1 1 0 1 1 1
21 0 0 0 0 1 0 1 0 1 1 1 1 1 1 1 1 1 1 1 1 1 0 1 1
22 0 0 0 1 1 1 1 0 1 0 1 1 1 1 1 1 1 0 1 1 1 1 0 1
23 0 0 0 0 1 0 1 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0

Table 1: Con ict Graph in matrix form

1000
Bidirectional MAC Houses

750
0.8
Throughput bound

0.6
Meters

500

0.4

250
0.2

Upper Bound
Lower Bound
0
0 20 40 60 80 100 0
0 250 500 750 1000
Effort Meters

Figure 5: Throughput with a bidirectional MAC Figure 6: Neighborhood Map

We start with several simplifying assumptions. We will

relax these assumptions as we proceed through the section. put is 0.5. A convenient way to visualize the optimal trans-
We assume that the range of each node is one unit, i.e., just mission schedule is to imagine that time is divided into slots
enough to reach its lateral neighbors, but not the diagonal of equal size, and in each slot we can transmit one packet be-
ones. We also assume that the interference range is equal tween neighboring nodes, subject to constraints imposed by
to the communication range. We assume an 802.11-like pro- the con ict graph. Then, the following transmission sched-
tocol model of interference described in Section 3.5. This ule will achieve optimal throughput: (i) 0 ! 1 (ii) 1 ! 2
model requires both the sender and the receiver to be free (iii) 0 ! 3 and 2 ! 5 (iv) 3 ! 6 and 5 ! 8 (v) 0 ! 1
of interference for successful communication. We term this and 6 ! 7 (vi) : : : We can continue in this manner inde-
a bidirectional MAC. The resulting con ict graph for this nitely. It is easy to see that in alternate timeslots, node 0
scenario is shown in the matrix form in Table 1. A 0 indi- gets to transmit to either node 1 or 3. Hence the optimal
cates that the links are not in con ict with each other, while throughput is 0.5.
1 indicates otherwise. For example, when node 0 is trans- In Figure 5, we show the upper and lower bound on through-
mitting to node 3, node 1 can hear these transmissions, and put calculated by our model, as we devote increasing amount
hence can not transmit to node 2. Thus, links 1 (0 ! 1) of eort. As shown, the upper bound quickly converges to
and 3 (1 ! 2) are in con ict. the stable value of 0.667, which is somewhat higher than
We allow multipath routing. We assume that all wireless the optimal value. This is a clear indication of the fact that
links have an identical capacity (i.e., speed) of 1 unit and clique constraints alone are not suÆcient to guarantee opti-
that all nodes have innite buers. We designate node 0 to mality, even in such small graphs (as noted in Section 3.2.4).
be the sender, and node 8 to be the receiver. The sender The lower bound, on the other hand, steadily converges to
always has data to send, and the receiver is always willing the optimal value of 0.5. We have veried that our pro-
to consume the data. gram has discovered all independent sets and cliques with
In this scenario, it is easy to see that the optimal through- 100 units of eorts.
 Scenario Optimal Throughput
I 0.5
1000
House with wireless connection
Wireless links
Internet access point
Senders II 0.5
III 1
750
IV 1
Table 2: Throughputs for neighborhood mesh in
various scenarios
Meters

500

The results indicate that neither multipath routing nor

250 doubling the range of the radio increases cumulative through-
put in the scenario we considered. On the other hand, by
using two channels instead of one, the network may achieve
the maximum possible throughput of 1. The maximum pos-
0
sible throughput is 1 because the Internet access point has
only one radio. On the other hand, even if we use two radios,
0 250 500 750 1000
Meters

the throughput remains at one. It is not hard to see why.

Figure 7: Mesh formation in the neighborhood The situation is equivalent to having two separate copies of
the baseline network, and then adding up their throughputs.
These scenarios illustrate that the model we have developed
could be a useful, albeit rudimentary, tool for analysis and
Our model can also incorporate single path routing, mul- capacity planning in wireless multi-hop networks.
tiple source-destination pairs, multiple channels as well as
multiple radios. We demonstrate this exibility with a com- 4.2 Tradeoff between connectivity and through-
munity mesh networking scenario, in which multiple users put
share an Internet connection, using a multi-hop wireless net- In Section 3, we discussed how our model can accommo-
work. We consider a map of a real suburban neighborhood date nodes which do not send data in a greedy fashion, i.e.
shown in Figure 4.1. There are 252 houses in an area of 1 they have a lower send rate. In [17, 20], the authors have
square kilometer. We select 35 of these houses at random, shown that the per node capacity in the network decreases
and assume that these houses are equipped with hardware as the number of nodes in the network goes up. These re-
that enables them to participate in a wireless mesh network. sults, however, were derived under the assumption that each
We assume that communication range of the wireless tech- node sends data as fast as it can. In other words, the desired
nology is 200 meters, while the interference range is 400 sending rate of the node is assumed to be 1. However, if each
meters. In Figure 4.1, we show the resulting network (i.e., node has a lower desired sending rate, the richer connectivity
the connectivity graph). We select a node that is roughly provided by additional nodes might help increase per node
at the center of the area and designate it as the Internet ac- throughput, by allowing better routes to be discovered. We
cess point. We assume that there are four senders, located now explore this hypothesis using our model.
as shown in the Figure. All the senders communicate with We consider a 7x7 grid, whose nodes are 200 meters apart
the Internet access point, and the metric of interest is the horizontally, and vertically. We assume that the communi-
cumulative throughput of these senders. We assume that all cation range is 250 meters, and the interference range is 500
wireless links are of unit capacity. meters. We set the link capacity to 1. We assume a bidirec-
We begin with a baseline case, for which we assume a tional MAC, similar to the one used to plot Figure 5. We
bidirectional MAC and single path routing. Our linear pro- use single-path routing.
gram is set to optimize the sum of the throughputs of the We pick N nodes from the 49 available nodes, at random,
four ows, with no consideration for fairness. In this case, and without replacement. Half of these nodes are designated
with about 5000 units of eort, upper and lower bounds con- as senders, and the other half are designated as receivers.
verge, and our model indicates that the maximum possible The senders and the receivers form N=2 ows in the net-
cumulative throughput is 0.5. work. Each sender is paired with only one receiver. We
We may now ask what we can do to improve the cumula- rst calculate the fraction of ows for which the source and
tive throughput. We consider four possibilities: (I) Employ the destination lie in the same connected component of the
multi-path routing (II) Double the range of each radio. We topology. We call this fraction the connectivity ratio. The
also double the interference range. (III) Leave the radio connectivity ratio for various values of N is shown in Fig-
range unchanged, but use two non-overlapping channels in- ure 8. The results show that after 24 nodes (i.e. 12 ows)
stead of one. A node may communicate on only one of the are selected, the connectivity ratio becomes 1.
two channels at any given time, but may switch between We then assign a sending rate of D to each sender. Then,
channels as often as necessary. (IV) Use two radios instead using our model, we calculate the optimal throughput using
of one at each node. The radios are assumed to be tuned to single-path routing. We divide the cumulative throughput
two xed, non-overlapping channels, so a node may commu- by the number of ows (i.e. N=2) to obtain average per- ow
nicate on the two channels simultaneously. The throughput throughput, and normalize it further by dividing it by D.
bounds in each of the four scenarios are shown in Table 2. The resulting normalized per- ow throughput for various
In each case, the upper and the lower bounds converge to values of N and D is plotted in Figure 9.
the same value, which indicates that the solution is optimal. Note that when the sending rate is 0.01, the normalized
 Scheduling Optimal 802.11
Routing

Optimal Optimal Alternative

throughput routing scheme
1 (solve LP) (ns simulation)
Shortest-path State of art State of art
under optimal (ns simulation)
0.8
scheduling
(solve LP)
Connectivity ratio

0.6

0.4 Figure 10: Four scenarios.

0.2 per- ow throughput continues to rise even after the connec-

tivity has reached 1. This means that the richer connectivity
provided by additional nodes allows for newer routes, and
0 allows extra traÆc to be sent through the network. How-
4 8 12 16 20 24 28 32
ever, if each node sends at rate 1, the node might have little
Number of nodes
capacity left to forward traÆc from other nodes. Thus, the
Figure 8: Connectivity Ratio for 7x7 grid
average per- ow throughput peaks early (i.e the network is
saturated), and then declines slowly, as new nodes join the
network, but fail to transmit most of their desired traÆc.
For sending rate of 0.1, the results are between these two
cases.
We stress that these results have been derived by assuming
optimal routing, as well as optimal scheduling of packets.
In the next section, we further discuss the impact of these
two assumptions. We also note that we have used a simple
grid topology in this example, and more experiments with
complex topologies are needed to verify the generality of this
result.
4.3 Benefits of optimal routing in absence of
optimal scheduling
100 As shown in the previous sections, the optimal through-
D = 0.01
D = 0.1 put is achieved by selecting optimal routes and scheduling
D=1 the links on the routes appropriately. A natural question
to ask is how much performance improvement is due to the
Normalized per-flow throughput

80
optimal route selection, and how much is due to the opti-
mal scheduling. Motivated by this question, we empirically
60
examine four scenarios shown in Figure 10. They corre-
spond to (i) optimal routing with optimal scheduling, (ii)
shortest-path routing with optimal scheduling, (iii) optimal
routing under 802.11 MAC, and (iv) shortest-path routing
40

under 802.11 MAC. We rst brie y describe the approach

20 we use to derive throughput for each case, and then present
the results.
Given a network topology, we apply the algorithm de-
0 scribed in Section 3 to compute the optimal throughput un-
4 8 12 16 20 24 28 32 36 40
der single-path routing. This corresponds to scenario (i).
Number of nodes
To derive the performance of optimal routing under 802.11,
Figure 9: Normalized per- ow throughput
we run ns-2 [4] simulations. To ensure that the packets
follow the optimal routes, we specify the optimal routes
obtained in Scenario (i) as the static routes in ns-2. The
throughput numbers from these simulations correspond to
scenario (iii).
We then repeat our simulation using AODV [24], a stan-
dard shortest path routing protocol. The resulting through-
put corresponds to the performance of the scenario (iv).
To minimize the impact of AODV routing overhead, all numFlow run 1 run 2 run 3 run4 run 5
nodes are static and simulations are run for 50 seconds, long 2 1.00 1.25 1.6 1.3812 1.00
enough to make the initial route setup overhead negligible. 4 1.4092 1.0001 1.4369 1.4333 1.1429
Based on the AODV simulation results, we obtain a set 8 2.1023 1.0000 1.0451 1.1131 1.1194
of links that are used in the shortest paths between sources
and destinations. We then modify the LP formulation (from
Section 3) used to compute bounds on the optimal through- Table 3: Throughput ratios between optimal rout-
put by excluding all but those links that lie on one or more ing and shortest path routing, both under optimal
of the shortest paths. (We do so by setting the capacity of scheduling in a 7x7 grid.
such links to zero.) We solve the resulting LP, and obtain
the throughput for scenario (ii). numFlow run 1 run 2 run 3 run4 run 5
Our aim is to compare throughput in scenario (i) to through- 2 1.0768 2.4323 1.5278 1.7974 1.1863
put in scenario (ii). Similarly, we compare scenarios (iii) 4 1.0662 1.5374 0.7854 1.0181 1.5475
and (iv) against each other. Note that we do not compare 8 3.5474 1.2239 0.5011 1.1389 0.3905
the throughput obtained by solving the LP model with the
throughput obtained from ns-2 simulations.
We consider these four scenarios in a 7x7 grid (49 nodes). Table 4: Throughput ratios between \optimal"
The horizontal and vertical separation between adjacent nodes path routing and shortest path routing, both under
is 200 meters. We assume the communication range to be 802.11 MAC in a 7x7 grid.
250 meters, and the interference range to be 500 meters. All
other parameters are at their default settings in ns-2. For bound will be schedulable. This leads to the question: how
each simulation run, we randomly pick a few pairs of nodes do we decide when to stop looking for even tighter bounds?
as sources and destinations; the source sends packets to the Given that the con ict graph may have an arbitrarily com-
corresponding destination at a constant bit rate equal to the plex structure, we cannot wait until the upper and lower
wireless link capacity. bounds are within a small percentage of each other for this
Table 3 and Table 4 show the throughput ratios between may never happen. Even after all the cliques are found, the
optimal routing and shortest path routing, under optimal upper bound may stay well above the optimal feasible so-
scheduling (based on our LP model) and under the 802.11 lution. Thus, there is no easy way to decide when to stop
MAC (based on ns-2 simulations), respectively. In all cases, the calculations. The data we present next does indicate,
optimal routing yields comparable or better throughput than however, that convergence is quite good in many scenarios.
the shortest path routing when optimal scheduling is used.
The benet of optimal routing varies with the number of 4.5 Computational Costs
ows, as well as with the locations of communicating nodes. In Section 4.1, we mentioned that the eort metric pro-
For instance, when the two ows are far apart and do not vides only a rough indication of the computational costs of
interfere with each other, the optimal path achieves the nding the bounds. We now provide more data in this re-
same throughput as the shortest path (e.g., numFlow=2 and gard. Note that much of the data provided is for the MAT-
run=1, 5); when the two ows interfere with each other, the LAB [2] solver to which we had ready access; as noted below,
optimal path takes a detour, which results in reduced inter- the CPLEX [5] solver oered a a speedup of about 7X, al-
ference and hence higher throughput (e.g., the case of num- beit on a somewhat faster CPU. Unfortunately, we only had
Flow=2 and run= 2, 3, 4). For similar reasons, the optimal limited access to the CPLEX resource. So there is the po-
path outperforms the shortest path even under the 802.11 tential for signicant improvements over the computational
MAC (as evident from the numbers for the corresponding costs reported here.
runs in Table 4). In Table 5, we consider the relationship between the size
On the other hand, the optimal path routing does not of the network and the amount of time required to compute
always outperform the shortest path routing under 802.11 upper and lower bounds. The table shows the bounds com-
MAC; this is increasingly common when the number of ows puted after 150,000 units of eorts for several grid sizes, and
is higher. This occurs because as network load increases, it the time required to compute them. In each case, there is
is harder to nd paths that do not interfere with other ows a single ow in the network, with its source and destination
in the absence of optimal scheduling. nodes at diagonally opposite corners of the grid. The rest
The above results are encouraging, and suggest that there of the parameters are similar to those used to plot Figure 5.
is a potential to improve throughput by making route selec- Note that the upper and lower bounds are not equal in all
tion interference-aware. In ongoing work, we are continuing cases, which indicates that we might not have found the
to investigate the benets of interference-aware routing un-
der a wider range of scenarios.
Grid Size Upper Bound Lower Bound Time (minutes)
4.4 Convergence of upper and lower bounds 3x3 0.25 0.25 2
In most of the previous results in this section, the upper 5x5 0.5 0.5 2
and the lower bounds converged, assuring us of the opti- 7x7 0.495 0.5 25
mality of the solution. When they did not converge, e.g., 9x9 0.474 0.5 35
Figure 5, we were able to assure ourselves of optimality of 11x11 0.479 0.5 40
the lower bound by manual verication. In general, how-
ever, the bounds may not converge, as there is no guarantee Table 5: Upper and lower bounds after 150,000 units
that even after adding all the clique constraints the upper of eort
 Eort Upper Bound Lower Bound Time (minutes) numbers for CPLEX1 .
10000 0.443 0.5 2 Since these numbers are based on a single run, and are
50000 0.48 0.5 5 based only on grid graphs, which have a regular connectiv-
100000 0.49 0.5 13 ity pattern, we cannot draw general conclusions from them.
150000 0.495 0.5 25 However, some trends are useful to note. We observe that
200000 0.5 0.5 41 for grid networks, the amount of time required to solve the
problem increases with the number of nodes. We also see
Table 6: Upper and lower bounds after varying ef- that for a given eort level, the time required to compute
fort for a 7x7 grid the bounds does not depend signicantly on the number of
ows in the network. However, the dierence between the
Flows Upper Bound Lower Bound Time (minutes) upper and lower bounds for a given amount of eort tends
2 0.578 0.583 34 to increase with increase in number of ows.
3 0.707 0.75 31 In case of irregular graphs, such as the neighborhood graph
4 0.758 0.833 29 shown in Figure 4.1, we have observed that the amount of
5 0.799 0.875 31 time required to solve the depends signicantly on connec-
tivity and interference patterns.
6 0.849 0.925 34 Finally, we note that we have not included any results
7 0.861 1.00 36 involving physical model of communication in this section.
We have also not included results that demonstrate the use
Table 7: 7x7 grid, multiple ows, 150,000 units of of links of dierent capacities. While we have solved such
eort networks (physical models of interference, links of dierent
capacities etc.), we could not do a detailed study due to
resource constraints. Therefore, we have chosen to focus on
optimal solution in all cases. The computations were done the protocol model of interference in this paper.
using MATLAB 6.1 [2], on a machine with 1.7Ghz Penti-
mum processor, and 1.7GB of RAM.
In Table 6, we consider the relationship between the amount 5. CONCLUSION AND FUTURE WORK
of eort, and the closeness of upper and lower bounds, as In this paper we have presented a model and methodology
well as the time required to compute those bounds. The for computing bounds on the optimal throughput that can
results are based on the 7x7 grid, with rest of the param- be supported by a multi-hop wireless network. A key dis-
eters similar to those used for Table 5. As we discussed tinction compared to previous work is that we work with any
in Section 4.1, with more eort, we are likely to add more given wireless network conguration and workload specied
variables as well as more restrictive constraints in the linear as inputs. No assumptions are made on the homogeneity
program. So the bounds become tighter. of nodes with regard to radio range or other characteristics,
In Table 7, we consider the relationship between the num- or regularity in communication pattern. We use a con ict
ber of ows in the network, and the amount of time required graph to model wireless interference under various conditions
to compute bounds for a given amount of eort. The results (multiple radios, multiple channels, etc.). We view the gen-
are based on a 7x7 grid, with multiple ows. For each ow, erality of our methodology and the con ict graph framework
the source is in the bottom row of the grid, and it commu- as a key contribution of our work.
nicates with a destination located in the same column, but Although the bounds that we compute on the optimal
in the top row. All other parameters are same as Table 5. throughput assume the ability to nely control and carefully
The software used to solve the linear program is also a schedule packet transmissions, the optimal routes yielded by
signicant factor in the amount of time required to nd our analysis often outperform shortest path routes even un-
the optimal solution. In Table 8, we show the amount of der \real-world" conditions such as uncoordinated schedul-
time taken by CPLEX [5] to solve the 7x7 grid case, with 6 ing and MAC contention. In ns-2 simulations, we have ob-
and 7 ows on a 2.7GHz Pentium machine, with 3.7GB of served a throughput improvement of over a factor of 2 in
RAM. While we can not compare these entries directly with some cases. The reason for this signicant improvement is
the corresponding entries in Table 8, as the machines used that the optimal routes often tend to be less interference-
to run MATLAB and CPLEX are dierent, the speedup is prone than the default shortest path routes.
still quite signicant: a reduction by a factor of 7, from We have also considered the impact of new nodes on the
34-36 minutes down to 5 minutes. Moreover, MATLAB per-node throughput in multi-hop wireless networks. Con-
cannot solve the Mixed Integer Programs that result when trary to previous results, we have found that the addition
single-path routing is used. We could only solve these using of new nodes can be benecial for all nodes, under the
CPLEX. Unfortunately, we only had limited access to the (realistic) assumption that each node is active for only a
CPLEX software, so we are unable to report the full set of small fraction of the time. The richer connectivity (and
hence increased opportunities for routing around interfer-
ence \hotspots" in the network) contributed by new nodes
Flows Upper Bound Lower Bound Time (minutes) more than oset the increase in traÆc load they cause.
6 0.849 0.925 5 In ongoing work, we are continuing to investigate the ben-
7 0.861 1.00 5 ets of interference-aware routing under a wide range of sce-
narios. Our next step after that would be to design a prac-
Table 8: 7x7 grid, multiple ows, 150,000 units of 1
Although we expect to be able to do so for the nal version
eort, with CPLEX of this paper should it be accepted for publication.
tical interference-aware routing protocol, which presents in- [21] M.Chudnovsky, Robertson, N., P.D.Seymour,
teresting challenges such as constructing the con ict graph and R.Thomas . The Strong Perfect Graph Theorem.
and computing optimal routes in a distributed manner. Submitted for publication. URL
http://www.math.gatech.edu/ethomas/spgc.html,
6. REFERENCES February 2003.
[1] Bay area wireless users group. [22] Park, V. D., and Corson, M. S. A highly adaptive
http://www.bawug.org/. distributed routing algorithm for mobile wireless
[2] Matlab version 6.1. http://www.matlab.com/. networks. In Proc. of IEEE INFOCOM'97 (Apr.
1997).
[3] Seattle wireless. http://www.seattlewireless.net/. [23] Highly dynamic
[4] Ns (network simulator), 1995. URL
Perkins, C. E., and Bhagwat, P.
destination-sequenced distance vector routing (dsdv)
http://www-mash.cs.berkeley.edu/ns/. for mobile computeres. In Proc. of ACM
[5] Ilog cplex suite, 2003. URL SIGCOMM'94 (Sep. 1994).
http://www.ilog.com/products/cplex/. [24] Perkins, C. E., and Royer, E. M. Ad-hoc
[6] Abramson, N. The aloha system | another on-demand distance vector routing. In Proc. of IEEE
alternative for computer communications. Proc. WMCSA'99 (Feb. 1999).
AFIPS Fall Joint Computer Conference (1970), [25] Priority scheduling in
281{285.
Yang, X., and Vaidya, N. H.
wireless ad hoc networks. In ACM MobiHoc (June
[7] Berkelaar, M. lp solve: linear programming code. 2002).
ftp://ftp.ics.ele.tue.nl/pub/lp solve/.
[8] ChavtaL, V. Linear Programming. W. H. Freeman APPENDIX
and Compnay, 1983.
[9] Couto, D. S. J. D., Aguayo, D., Chambers,
.1 Proof of Theorem 1
B. A., and Morris, R. Performance of multihop Suppose we are given a graph G and we want to compute
wireless networks: Shortest path is not enough. In 1st the cardinality of its maximum independent set. We now
Workshop on Hot Topics in Networks (Oct. 2002). construct a wireless network such that the optimal through-
[10] Estrin, D., Govindan, R., Heidemann, J., and put it can support under the protocol interference model is
Kumar, S. Next century challenges: Scalable same as the cardinality of the maximum independent set of
coordination in sensor networks. In ACM MOBICOM G. Create two wireless nodes, a source s and a receiver r.
(Aug. 1999). For every vertex in G add a wireless link of unit capacity
[11] Computers and between s and r. For every edge between two nodes in G,
assume a con ict between the corresponding wireless links
Garey, M. R., and Johnson, D. S.
Intracctability: A guide to the theory of NP
completness. W. H. Freeman and Company, 1979. in the network. (Such a network may arise, for instance, if
[12] On the capacity of nodes s and r are each equipped with multiple radios set
Gastpar, M., and Vetterli, M.
wireless networks: the relay case. In IEEE INFOCOM either to the same (i.e., interfering) channel or to separate
(Jun. 2002). (i.e., non-interfering) channels.) It is not hard to see that
[13] Mobility increases that the optimal throughput is achieved if and only if a max-
Grossglauser, M., and Tse, D.
the capacity of ad-hoc wireless networks. In IEEE imum independent set in G is scheduled. Thus nding the
INFOCOM (Apr. 2001). optimal capacity of the wireless network is equivalent to to
[14] nding the cardinality of the maximum independent set of
Grotschel, M., Lovasz, L., and Schrijver, A.
The ellipsoid method and its consequences in graph G, which is known to be a hard problem.
combinatorial optimization. Combinatorica 1 (1981), The above proof may come across as contrived since the
169{197. wireless network we constructed is unlikely to arise in prac-
[15] tice. This raises an interesting question of whether realistic
Grotschel, M., Lovasz, L., and Schrijver, A.
Geometric Methods in Combinatorial Optimization.
wireless networks could give rise to complex con ict graphs?
Academic Press, 1984. Our answer is both yes and no. Our answer is "yes" be-
cause the maximum independent set problem is hard due to
[16] Gupta, P., and Kumar, P. R. The capacity of the existence of odd holes and odd anti-holes in the given
wireless networks. IEEE Transactions on Information graph2 . As shown in Figure 11, very realistic and simple
Theory 46, 2 (Mar. 2000). grid graphs could have con ict graphs with many odd holes
[17] Gupta, P., and Kumar, P. R. The capacity of and odd anti-holes. On the other hand, our answer is "no"
wireless networks. IEEE Transactions on Information because realistic con ict graphs may have some special prop-
Theory 46, 2 (Mar. 2000), 388{404. erty or structure that could make the problem of nding the
[18] Johnson, D. B., and Maltz, D. A. Dynamic source maximum independent set easy. We have been unable to
routing in ad-hoc wireless networks. In Mobile identify any such property, but our failure does not mean
Computing (1996), T. Imielinski and H. Korth, Eds., that no such property exists (though the complex con ict
Kluwer Academic Publishers. graphs arising from the simple grid graphs, as in Figure 11,
[19] Khachiyan, L. G. A polynomial algorithm in linear diminish our optimism). In view of this, we believe that the
programming. Soviet Mathematics Doklady 20 (1979), heuristic approach presented in Section 3 may be reasonable.
191{194. 2
If a graph does not have any odd holes or anti-holes then
[20] Li, J., Blake, C., Couto, D. S. J. D., Lee, H. I.,
the graph is termed perfect [21], and for perfect graphs there
and Morris, R. Capacity of ad hoc wireless are polynomial time algorithms to solve the maximum inde-
networks. In ACM MOBICOM (Jul. 2001). pendent set problem [14].
 A
algorithm for nding the heaviest clique [15]. However, both
the ellipsoid algorithm 3and the semidenite algorithm have
W
a running time of O(n ), so in combination their running
E B
time is O(n6 ). Thus this polynomial time algorithm is not
very practical, which makes our heuristic approach attrac-
X tive even in this special case.
.3 Finding Violated Odd Hole Constraints
V
Now we present a separation oracle that given a con ict
D C graph G and a candidate solution  nds a violated odd hole
constraint, if any. Such an oracle could be used to improve
Y
the convergence rate of the algorithm presented in Section
3. Note that this separation oracle is applicable to general
graphs; for the perfect con ict graph considered in Section .2
above, there are no odd holes anyway.
Consider an odd hole, H , of the given con ict graph G.
Z

Any vector  inside thePindependent set polytope of G must

Figure 11: A 6x6 grid connectivity graph. ABCDE satisfy the following: 2   (jH j 1)=2. A violated i

odd hole is one for which this constraint is not satised.
i H

and VWXYZ are examples of odd holes in the cor- Before attempting to nd a violated odd hole, we may as-
responding con ict graph, assuming an 802.11-style sume that the given  satises all the edge constraints, i.e.,
MAC, communication range equal to the lateral spac-  +   1 for every edge in G, because if it does not then
ing between neighbors, and interference range equal to i j

we can include the violated edge constraint to shrink the

twice the communication range. These odd holes also upperbounding polytope. After making this assumption we
happen to be odd anti-holes. dene a weight function on the edges. For every edge ij of
the graph G, we dene its weight to be 1   , which i j

is guaranteed to be non-negative. With this weight func-

.2 Polynomial Time Algorithm in Special Case tion we nd the lightest (i.e., least-weight) odd cycle in the
Even in special cases where polynomial time algorithms graph. The lightest odd cycle can be found using a bipartite
may exist, they may be too expensive to be of practical in- graph construct as explained
be the lightest odd cycle.
P in2 the(1 next paragraph. Let C
  ) < 1 is equiv-
terest. One such special case arises in the context of grid
graphs when the con ict radius is zero. By zero con ict ra-
P
alent to 2  > j j2 1 . So, if the weight of the lightest
i
C
ij C i j

dius we mean that two links con ict if and only if they share odd cycle is less than 1 then the cycle is a violated odd hole.
i C

an endpoint. In this simple and somewhat unrealistic set- If the weight of the lightest odd cycle is 1 or more then there
ting, the con ict graph is nothing but the line graph of the is no violated odd hole.
underlying grid network. (The line graph, L(G), of a graph, Now we come to the question of eÆciently nding the
G, is a graph on the edges of G, i.e., the vertices of L(G) lightest odd cycle. Let G be the graph in which we need to
correspond to the edges of G. There is an edge between nd the lightest odd cycle. We construct a bipartite graph,
two vertices of L(G) if the corresponding edges in G have a B , as follows. For every vertex v in G we put two vertices
vertex in common.) Our network in this case is a grid. A v and v in B (the subscripts l and r can conceptually be
l r

grid is a bipartite graph, and bipartite graphs are perfect. thought of as representing the left and right \halves" of the
The line graph of a perfect graph is perfect too. Hence the the bipartite graph B ). For every edge uv in G we put two
con ict graph of a grid graph with a zero con ict radius is a edges u v and u v in B . Now an odd cycle in G becomes
l r r l

perfect graph. A perfect graph has the property that its set an odd length path in B e.g., uvwu becomes u v w u . So l r l r

of clique constraints dene its independent set polytope. So for every vertex u in G we nd the shortest path from u to l

if we write a linear program with all the clique constraints u in B . The shortest such path in B yields the lightest odd
r

together with the ow constraints then we can nd the opti- cycle in G.

mal network throughput. The problem, however, is that the
number of cliques could still be exponentially many. (Al-
though this does not happen with grid graphs, it could very
well happen with other perfect graphs.) A solution is to
use the ellipsoid algorithm [19] to optimize linear functions
over a polytope. This algorithm does not require all the
constraints in an explicit form to optimize a linear function
over a polytope, hence we do not have to enumerate the
exponentially many clique constraints. The ellipsoid algo-
rithm only needs a subroutine that given a potential solu-
tion indicates whether the constraints are satised or not,
and if not identies at least one constraint which is not sat-
ised. Such a subroutine is called separation oracle. The
separation oracle for our problem would be one that nds a
violated clique constraint given a usage vector. This can be
accomplished using the Grotschel semidenite programming