Anda di halaman 1dari 5

5822 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO.

12, DECEMBER 2009

Optimal Spectrum-Efficient Routing in Multihop Wireless Networks


Mohamed Saad, Member, IEEE

Abstract—This paper addresses the problem of finding the Studies on routing that combine tools from information
route with maximum end-to-end spectral efficiency, under the theory and networking do exist. See, e.g., [2] and [10]. In
constraint of equal bandwidth sharing, in multihop wireless particular, [10] considers joint routing and resource allocation
networks. This problem has been addressed recently in the litera-
ture, and only exhaustive search with exponential computational in wireless multihop networks, such that a network-wide
complexity or suboptimal heuristics are known. This paper closes objective (e.g., delay) is optimized. The recent study in [2]
the algorithmic gap by introducing two algorithms that provide introduces the problem of finding the path with maximum
provably optimal solutions to the problem in polynomial-time. spectral efficiency in multihop wireless networks, under the
The proposed algorithms rely on the iterative use of a shortest constraint of equal bandwidth sharing. On the one hand, the
path procedure. Our computational results further illustrate the
efficiency of the proposed approach. authors of [2] note that simple shortest path algorithms cannot
be used to solve the problem because the resulting routing
Index Terms—Multihop wireless networks, spectrum-efficient metric is neither isotonic nor monotone [8]. On the other hand,
routing, polynomial-time algorithms, divide-and-conquer.
exhaustive search has an exponential computational complex-
ity because it involves pre-computing all paths joining a given
I. I NTRODUCTION node pair. Therefore, the study in [2] proposes two efficient,
yet sub-optimal spectrum-efficient routing heuristics. The goal

M ULTIHOP wireless networks consist of a set of


wireless devices that communicate with each other
over multiple wireless hops, with participating nodes col-
of this paper, however, is to study efficient (polynomial-time)
algorithms that are able to provide exact optimal solutions to
the problem.
laboratively forwarding ongoing traffic. Wireless multihop The contribution of this paper can be summarized as fol-
relaying/forwarding is the foundation for the development lows.
and deployment of emerging technologies such as in- ∙ We close the algorithmic gap by showing (in Section III)
frastructure wireless mesh networks [1] (where wireless that a provably optimal solution to the spectrum-efficient
routers/access points are interconnected to provide an infras- routing problem can be obtained in polynomial-time. In
tructure/backbone for clients) or IEEE 802.16j relay-based particular, we show that, using the divide-and-conquer
networks [6] (where a number of fixed relays can be used principle, a provably optimal solution to the problem
to extend the coverage area of a base station, and/or increase can be obtained by iteratively solving a sequence of
the capacity of a wireless access system). relaxations. This leads to a polynomial-time algorithm
The study of wireless multihop relaying (including routing that relies entirely on the iterative use of a shortest path
as a special case) has received considerable attention in procedure.
two research communities: the information theory community ∙ We present (in Section IV) a technique that further re-
and the networking community. The foremost contribution of duces the average complexity of the proposed algorithm.
information-theoretic studies (see, e.g., [7] and [11]) is the Our computational results (in Section V) illustrate the
understanding of the fundamental limits on network perfor- efficiency of the proposed approach.
mance measures, e.g., transport capacity. These information-
theoretic studies have led to several relaying protocols that II. P ROBLEM D EFINITION
are order-optimal as the number of nodes goes to infinity.
A multihop wireless network is modeled as a graph 𝐺 =
These protocols, however, are typically too complex to be
(𝑉, 𝐸), where 𝑉 represents the set of nodes (vertices) and 𝐸
implemented in practical systems [2].
represents the set of links (edges). We let 𝑙 ∈ 𝐸 signify a link
Studies within the networking community on wireless mul-
in the network. We also let 𝑁 = ∣𝑉 ∣ and 𝑀 = ∣𝐸∣ denote the
tihop routing focus mainly on the design of new routing
number of nodes and links, respectively.
metrics, and the modification of existing routing protocols
Following [2], we consider the setting in which all transmit
to incorporate new metrics (see, e.g., [3] and [5]). These
devices are constrained by the same symbol-wise average
studies, however, are often built on link-level abstractions
transmit power 𝑃 , and assume that all devices transmit with
of the network without fully considering the impact of the
power 𝑃 when transmitting. A possible justification for this
physical layer [2].
assumption is that nodes in infrastructure wireless mesh
networks are mostly immobile and connected with abundant
Manuscript received April 19, 2009; revised August 22, 2009; accepted
October 19, 2009. The associate editor coordinating the review of this letter power supplies. Therefore, for a link 𝑙 ∈ 𝐸, the signal-to-noise
and approving it for publication was F.-N. Pavlidou. ratio (SNR) is given by:
The author is with the Department of Electrical and Computer Engineering,
University of Sharjah, Sharjah, UAE (e-mail: msaad@sharjah.ac.ae). 𝑃 𝐺𝑙
Digital Object Identifier 10.1109/TWC.2009.12.090546
𝑆𝑁 𝑅𝑙 = , (1)
𝑁0 𝐵
1536-1276/09$25.00 ⃝
c 2009 IEEE
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO. 12, DECEMBER 2009 5823

where 𝐺𝑙 is the path gain from the sender of link 𝑙 to the solved by searching over all possible values of 𝛼(𝐿). In
receiver of link 𝑙, 𝑁0 is the normalized one-sided power particular, (5) can be solved using the following procedure:
spectral density of the additive white Gaussian noise (at any
receiver in the network), and 𝐵 is the finite bandwidth of the Procedure 1: { }
wireless channel. ∙ For every 𝑎 ∈ log(1 +
𝑃 𝐺𝑙
𝑁0 𝐵 ) : 𝑙 ∈ 𝐸 , solve:
The spectral efficiency 𝑅(𝐿) of an arbitrary path 𝐿 in the
network is defined as the bandwidth-normalized end-to-end 𝛼(𝐿)
rate, i.e., 𝑅(𝐿) = 𝐶𝐿 /𝐵 (in bps/Hz), where 𝐶𝐿 is the end-to- max . (6)
𝐿∈ℒ𝑠𝑑 :𝛼(𝐿)=𝑎 ∣𝐿∣
end achievable rate (in bps) and 𝐵 is the channel bandwidth (in
Remark 1. Let path 𝐿∗ be the optimal solution to (5),
Hz) [2]. Following [2], it is assumed that a common channel
and let 𝑅(𝐿∗ ) denote its spectral efficiency, i.e., 𝑅(𝐿∗ ) is
is shared among all nodes using time division multiple access
the optimal
{ objective function} value of (5). Also, for a given
(TDMA) without spatial reuse, i.e., each node transmits in 𝑃 𝐺𝑙
its own unique time slot (justifications for this assumption 𝑎 ∈ log(1 + 𝑁 0𝐵
) : 𝑙 ∈ 𝐸 , let path 𝐿𝑎 be the optimal
can be found in Sections II.C and VI of [2]). The medium solution to (6), and let 𝑅(𝐿𝑎 ) denote its spectral efficiency. In
access control (MAC) layer of the IEEE 802.16 mesh protocol, other words, 𝑅(𝐿𝑎 ) is the optimal objective function value of
for example, is based on TDMA (see, e.g., [4]). Under the (6). Since the union of the{sets {𝐿 ∈ ℒ𝑠𝑑 : 𝛼(𝐿) } = 𝑎} across
𝑃 𝐺𝑙
constraint of equal bandwidth sharing, the end-to-end spectral all possible values of 𝑎 ∈ log(1 + 𝑁 0𝐵
) : 𝑙 ∈ 𝐸 covers the
efficiency 𝑅(𝐿) of a path 𝐿 is given by [2] set of paths ℒ𝑠𝑑 , and by the divide and conquer principle [9],
1 𝑃 𝐺𝑙 the following is true:
𝑅(𝐿) = min log(1 + ), (2)
𝑙∈𝐿 ∣𝐿∣ 𝑁0 𝐵 𝑅(𝐿∗ ) = max 𝑅(𝐿𝑎 ), (7)
𝑎∈𝐴
where ∣𝐿∣ is the number of links (hops) in path 𝐿. Note that { }
𝑃 𝐺𝑙
the factor 1/∣𝐿∣ results from the sharing of bandwidth equally where the set 𝐴 = log(1 + 𝑁 0𝐵
) : 𝑙 ∈ 𝐸 .
among relay links [2]. Moreover, by substituting 𝛼(𝐿) = 𝑎 in the objective
For a given path 𝐿, we define function, it is easily seen that (6) is equivalent to the following
𝑃 𝐺𝑙 problem:
𝛼(𝐿) = min log(1 + ). (3) 𝑎
𝑙∈𝐿 𝑁0 𝐵 max . (8)
𝐿∈ℒ𝑠𝑑 :𝛼(𝐿)=𝑎 ∣𝐿∣
By combining (2) and (3), the end-to-end spectral efficiency
Instead of solving (8) directly, we consider its relaxation
of any path 𝐿 is given by
in which 𝛼(𝐿) = 𝑎 is replaced with the relaxed constraint
𝛼(𝐿) 𝛼(𝐿) ≥ 𝑎. Now, consider the following modified procedure
𝑅(𝐿) = . (4)
∣𝐿∣ to solve (5):
Given a source-destination (𝑠-𝑑) pair of nodes (𝑠, 𝑑) ∈ 𝑉 × 𝑉 , Procedure 2: { }
the problem of finding the route with maximum end-to-end 𝑃𝐺
∙ For every 𝑎 ∈ log(1 + 𝑁 𝐵𝑙 ) : 𝑙 ∈ 𝐸 , solve:
0
spectral efficiency under the constraint of equal bandwidth
𝑎
sharing can, thus, be expressed as the following optimization max . (9)
𝐿∈ℒ𝑠𝑑 :𝛼(𝐿)≥𝑎 ∣𝐿∣
problem:
𝛼(𝐿) Problem (9) is a relaxation of (8) in the sense that any
max , (5)
𝐿∈ℒ𝑠𝑑 ∣𝐿∣ feasible solution (path) for (8) is also feasible for (9).
Remark 2. Let the paths 𝐿𝑎 and 𝐿 ˆ 𝑎 be the optimal solutions
where ℒ𝑠𝑑 is the set of all paths 𝐿 connecting the source-
destination pair of nodes (𝑠, 𝑑) ∈ 𝑉 × 𝑉 . to problems (8) and (9), respectively. Since 𝐿𝑎 is optimal for
(8), it is also feasible for (9). Therefore, 𝑎/∣𝐿𝑎 ∣ ≤ 𝑎/∣𝐿 ˆ 𝑎∣ .
As noted in [2], Bellman-Ford or Dijkstra shortest path
algorithms cannot be used to solve (5) directly because the This implies that
ˆ 𝑎 ∣ ≤ ∣𝐿𝑎 ∣.
∣𝐿 (10)
routing metric 𝛼(𝐿)/∣𝐿∣ is neither isotonic nor monotone [8].
Problem (5), however, can be solved by an exhaustive search The main result follows. { }
over all (possibly exponentially many) paths. This exponential Proposition 1: For a given 𝑎 ∈ log(1 + 𝑁 𝑃 𝐺𝑙
) : 𝑙 ∈ 𝐸 ,
0𝐵
complexity makes exhaustive search prohibitive in networks ˆ 𝑎 denote the optimal solution to (9), and let 𝑅(𝐿 ˆ 𝑎) =
with a moderate-to-large number of nodes. The study in let path 𝐿
𝛼(𝐿ˆ𝑎)
[2] presents two efficient algorithms that provide suboptimal ˆ𝑎∣
∣𝐿
denote its spectral efficiency. Also, let path 𝐿∗ be the
solutions to (5). In what follows, however, we show that 𝛼(𝐿∗ )
optimal solution to (5), and let 𝑅(𝐿∗ ) = ∣𝐿∗ ∣ denote its
provably optimal solutions to the problem can be obtained spectral efficiency. Then
in polynomial-time.
𝑅(𝐿∗ ) = max 𝑅(𝐿 ˆ 𝑎 ), (11)
𝑎∈𝐴
III. O PTIMAL P OLYNOMIAL -T IME A LGORITHM { }
𝑃 𝐺𝑙
The main idea of the proposed algorithm is the fact that, where 𝐴 = log(1 + 𝑁 0𝐵
) : 𝑙 ∈ 𝐸 .
for any path 𝐿, 𝛼(𝐿) takes one of finite possible values. In Proof: The following is true:
particular,{ it is easily seen from } the definition in (3) that 𝛼(𝐿ˆ𝑎) 𝑎
𝑃 𝐺𝑙 ˆ 𝑎) =
𝑅(𝐿 ≥ , (12)
𝛼(𝐿) ∈ log(1 + 𝑁 0𝐵
) : 𝑙 ∈ 𝐸 . Consequently, (5) can be ˆ 𝑎∣
∣𝐿 ˆ 𝑎∣
∣𝐿
5824 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO. 12, DECEMBER 2009

where the inequality follows from the fact that 𝛼(𝐿 ˆ𝑎) ≥ 𝑎 IV. R EDUCED C OMPLEXITY A LGORITHM
ˆ
because path 𝐿𝑎 is the optimal solution to (9). Now, let path In this section, we explore a possibility for reducing the
𝐿𝑎 denote the optimal solution for problem (8), and let 𝑅(𝐿𝑎 ) complexity of our proposed Algorithm A. We begin with the
denote its spectral efficiency. Thus, following observation.
𝑎 𝑎 Proposition 2: Let path 𝐿𝑠 denote the minimum-hop path
𝑅(𝐿𝑎 ) = ≤ ˆ 𝑎 ).
≤ 𝑅(𝐿 (13)
∣𝐿𝑎 ∣ ˆ 𝑎∣
∣𝐿 in the network from source 𝑠 to destination 𝑑. Also, let path
𝐿∗ be the optimal solution to (5), i.e., the path with maximum
Note that the equality follows from the fact that 𝛼(𝐿𝑎 ) = 𝑎 spectral efficiency. Then 𝛼(𝐿∗ ) ≥ 𝛼(𝐿𝑠 ).
because 𝐿𝑎 is optimal for (8). The first inequality follows from Proof: Since 𝐿∗ is the path with maximum spectral
Remark 2, and the second inequality follows from (12). efficiency, then 𝑅(𝐿∗ ) ≥ 𝑅(𝐿𝑠 ). Equivalently,
Combining Remark 1 and (13), leads to: 𝛼(𝐿∗ ) 𝛼(𝐿𝑠 )
≥ . (15)
ˆ 𝑎 ) ≥ 𝑅(𝐿∗ ), ∣𝐿∗ ∣ ∣𝐿𝑠 ∣
max 𝑅(𝐿 (14)
𝑎∈𝐴
Moreover, 𝐿𝑠 is the path with the minimum number of hops.
{ } Therefore,
𝑃 𝐺𝑙
where 𝐴 = log(1 + 𝑁 ) : 𝑙 ∈ 𝐸 . Note that among all

0𝐵 ∣𝐿𝑠 ∣ ≤ ∣𝐿∗ ∣. (16)
paths 𝐿 ∈ ℒ𝑠𝑑 , 𝑅(𝐿 ) is the maximum possible spectral
efficiency. For any 𝑎 ∈ 𝐴, 𝐿 ˆ 𝑎 ∈ ℒ𝑠𝑑 , i.e., 𝐿
ˆ 𝑎 is a feasible Combining (15) and (16) completes the proof.
path for problem (5). Therefore, (14) must hold with strict In light of Proposition 2, we need to search only over values
equality. of 𝛼(𝐿) that are at least as large as 𝛼(𝐿𝑠 ), with 𝐿𝑠 being the
Now, the only remaining issue is to show how to solve (9). minimum-hop path between the given 𝑠-𝑑 pair. Algorithm A
For any given value for 𝑎, maximizing 𝑎/∣𝐿∣ is equivalent to can, thus, be modified as follows:
minimizing ∣𝐿∣. Consequently, problem (9) is equivalent to Algorithm B: Modified Optimal
finding the minimum-hop path 𝐿 from source 𝑠 to destination Spectrum-Efficient Routing:
𝑑, such that 𝛼(𝐿) ≥ 𝑎. It can be seen from (3) that the value ∙ Find 𝐿𝑠 , the minimum-hop path from
of 𝛼(𝐿) for any path 𝐿 is determined by its bottleneck link. source 𝑠 to destination { 𝑑. }
Therefore, problem (9) can be solved, for a given 𝑠-𝑑 pair and Let 𝛼(𝐿𝑠 ) = min𝑙∈𝐿𝑠 log(1 + 𝑁 𝑃 𝐺𝑙
) , and let
0𝐵
a given value for 𝑎, as follows: 𝛼(𝐿𝑠 )
𝑅(𝐿𝑠 ) = ∣𝐿𝑠 ∣ .

𝑃 𝐺𝑙
Remove all links 𝑙 ∈ 𝐸 for which log(1 + 𝑁 ) < 𝑎. In { }
0𝐵 𝑃 𝐺𝑙
the remaining graph, obtain the minimum-hop path using ∙ For every 𝑎 ∈ log(1 + 𝑁0 𝐵 ) : 𝑙 ∈ 𝐸 , do:
Dijkstra or Bellman-Ford algorithms. – If 𝑎 = 𝛼(𝐿𝑠 ) then 𝑅(𝐿 ˆ 𝑎 ) = 𝑅(𝐿𝑠 ).
In light of the above discussion, Procedure 2 becomes – If 𝑎 > 𝛼(𝐿𝑠 ), do
equivalent to the following algorithm: ∗ Remove all links 𝑙 ∈ 𝐸 for which
𝑃 𝐺𝑙
Algorithm A: Optimal Spectrum-Efficient log(1 + 𝑁 0𝐵
) < 𝑎.
Routing: ∗ In the remaining graph, find 𝐿 ˆ 𝑎,
{ }
𝑃𝐺 the minimum-hop path from 𝑠 to
∙ For every 𝑎 ∈ log(1 + 𝑁 𝐵𝑙 ) : 𝑙 ∈ 𝐸 , do:
0
𝑑. Let 𝑅(𝐿 ˆ 𝑎 ) = 𝛼(𝐿ˆ 𝑎 ) .
ˆ ∣
∣𝐿
– Remove all links 𝑙 ∈ 𝐸 for which 𝑎

𝑃 𝐺𝑙 ∙ Return the path with largest 𝑅(𝐿 ˆ 𝑎 ).


log(1 + 𝑁 0𝐵
) < 𝑎.
– In the remaining graph, find 𝐿 ˆ 𝑎, Note that Algorithm B will only call a shortest path
the minimum-hop path from source 𝑠 procedure if 𝑎 ≥ 𝛼(𝐿𝑠 ). Therefore, the number of shortest-
ˆ 𝑎 ) = 𝛼(𝐿ˆ 𝑎 ) .
to destination 𝑑. Let 𝑅(𝐿 path iterations of Algorithm B is expected to be lower than
ˆ𝑎∣
∣𝐿 that of Algorithm A. It is easily seen, however, the worst-
∙ ˆ 𝑎 ).
Return the path with largest 𝑅(𝐿 case complexities of Algorithms A and B are both equal to
It was noted in [2] that (5) cannot be solved directly by a 𝑂(𝑁{4 ). This worst-case scenario
} occurs if all elements of the
𝑃 𝐺𝑙
shortest path algorithm. We have seen, however, that (5) can set log(1 + 𝑁0 𝐵 ) : 𝑙 ∈ 𝐸 are at least as large as 𝛼(𝐿𝑠 ).
be solved using an iterative use of a shortest path procedure. Our numerical results, however, illustrate that the average
Remark 3. By Proposition 1, and since Algorithm A and running time of Algorithm B is considerably less than that
Procedure 2 are equivalent, the path returned by Algorithm A of Algorithm A.
is an optimal solution for the original problem (5). Moreover, Note also that Algorithm B (respectively, Algorithm A) is
each iteration of the algorithm finds a minimum-hop path. If based entirely on an iterative use of a shortest path procedure.
Dijkstra algorithm is used, the computational complexity of Since each iteration finds a minimum-hop path, all link
each iteration is 𝑂(𝑁 2 ), where 𝑁 = ∣𝑉 ∣ is the number of distances are unity. To implement Algorithm B (respectively,
𝑃 𝐺𝑙
network nodes. The total number of iterations is 𝑀 , where Algorithm A), however, the value of log(1 + 𝑁 0𝐵
) has to be
𝑀 = ∣𝐸∣ is the number of network links. Consequently, the known for each link 𝑙. In practice, the link SNR (𝑃 𝐺𝑙 /𝑁0 𝐵)
overall complexity of the algorithm is 𝑂(𝑁 2 𝑀 ). Since an 𝑁 - can be directly measured by received signal strength indicators
node network has at most 𝑁 (𝑁 − 1) directed links, the worst available on most devices [2], and fed back to the transmitters.
case complexity of Algorithm A is 𝑂(𝑁 4 ). Nodes can then exchange their knowledge about the values
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO. 12, DECEMBER 2009 5825

0.15 0.09
Algorithm A

Average Algorithm Running Time (in seconds)


Algorithm B 0.08
Algorithm A
Algorithm Running Time (in seconds)

0.07 Algorithm B

0.1 0.06

0.05

0.04

0.05 0.03

0.02

0.01

0 0
0 20 40 60 80 100 −20 −15 −10 −5 0 5 10 15 20
Network Realization Network SNR (dB)

Fig. 1. Reduced running time of Algorithm B as compared to Algorithm A. Fig. 2. Reduced average running time of Algorithm B relative to Algorithm A
for different network SNRs. Every point in the figure is obtained by averaging
over 104 random (and independent) network realizations.
𝑃 𝐺𝑙
of log(1 + 𝑁0 𝐵 ) for their outgoing links using a link-state
1
protocol. 10
Algorithm A

Average Algorithm Running Time (in seconds)


Algorithm B

V. N UMERICAL R ESULTS
0
We consider multihop wireless networks, in which the nodes 10

are located at random positions in a 100 m × 100 m two-


dimensional area. It is assumed that there exists a link between
two nodes if the distance between the nodes is at most 25 m. −1
10
For each link 𝑙, it is assumed that the path gain is given by
−4
𝐺𝑙 = 𝑐 ⋅ 𝐴𝑙 ⋅ (𝑑𝑙 /𝑑0 ) , where 𝑑𝑙 is the length of link 𝑙, 𝑑0 is
the reference distance, 𝐴𝑙 is a log-normally distributed random
−2
variable (with 0-dB mean and 8-dB log-variance) that reflects 10

shadowing, and 𝑐 is a constant. Without loss of generality, we


set 𝑑0 = min𝑙 {𝑑𝑙 } and 𝑐 = 1/𝑚𝑎𝑥𝑙 {𝐴𝑙 }. The resulting path
gains have, thus, non-negative values smaller than one. We test −3
10
our proposed algorithms on random and independent network 10 20 30 40 50 60
Number of Nodes
realizations, where in each realization the horizontal and
vertical coordinates of each node are chosen randomly (and Fig. 3. Reduced average running time of Algorithm B relative to Algorithm
independently) according to a uniform distribution between 0 A for different numbers of nodes. Every point in the figure is obtained by
and 100 m, and the path gains are generated randomly (and averaging over 104 random (and independent) network realizations.
independently) according to the above-described model. More-
over, two arbitrary nodes are chosen as the 𝑠-𝑑 pair. All results
were obtained using Matlab implementations performed on a other words, Algorithm B reduced the average running time
1.86 GHz IntelⓇ Core2 processor and 1 GB of memory. by 90.77% as compared to Algorithm A.
First, we consider 20-node network realizations. Without Note also that, averaged over 104 additional random net-
loss of generality, the network SNR (𝑃/𝑁0 𝐵) is assumed to work realizations, the running time was 0.0982 seconds for
be 40 dB. We generate 100 random network realizations. For Algorithm A, and 0.0094 seconds for Algorithm B. In other
each realization, the path from 𝑠 to 𝑑 with maximum spectral words, Algorithm B reduced the average running time by
efficiency is chosen using both Algorithm A and Algorithm 90.43% relative to Algorithm A.
B. Our main focus is the spectral efficiency of the obtained Fig. 2 depicts the running time of Algorithm A and Algo-
paths, and the running time of the algorithms. As theoretically rithm B, respectively, when the network SNR (𝑃/𝑁0 𝐵) is
justified by Proposition 1 and Proposition 2, the spectral varied from -20 dB to 20 dB. In particular, for each network
efficiencies attained by Algorithm A and Algorithm B were SNR value, the running time of each algorithm is obtained
identical in every experiment. In particular, both algorithms by averaging over 104 random 20-node network realizations.
result in the provably optimal spectral efficiency. The running Algorithm B reduced the average time to obtain an exact
time results are depicted in Fig. 1. Obviously, the running time optimal path by 91.72% – 92.57% relative to Algorithm A.
of Algorithm B is consistently lower than that of Algorithm Fig. 3 depicts the running times of the algorithms when
A. In particular, the average running time was 0.0921 seconds the number of nodes 𝑁 is varied from 10 to 60. For each
for Algorithm A, and 0.0085 seconds for Algorithm B. In value for 𝑁 , the running time of each algorithm is obtained
5826 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO. 12, DECEMBER 2009

70
Relative Gap to Optimality (in %) Caused by DSER network realizations, the relative sub-optimality introduced by
DSER ranged from 0 to 83.9%, with an average of 27.68%.
60
The results signify the importance of our proposed algorithms
in assessing the sub-optimality caused by heuristics, especially
50
for a moderate-to-high number of nodes, i.e., when exhaustive
search is prohibitive.
40

VI. C ONCLUSION
30
This paper addressed the problem of finding the path with
20
maximum end-to-end spectral efficiency in multihop TDMA
wireless networks with equal bandwidth sharing. This problem
10
was introduced recently in the literature, and only exponential-
time exhaustive search or suboptimal heuristics were known.
0 We closed the algorithmic gap by introducing a polynomial-
0 20 40 60 80 100 time algorithm that provides provably optimal solutions to
Network Realization
the problem, while relying entirely on the iterative use of
Fig. 4. Relative gap to optimality (in %) caused by the distributed spectrum- a shortest path procedure. We also proposed an improved
efficient routing (DSER) algorithm introduced in [2]. version of the algorithm that reduced the average computation
time by more than 90% in most numerical experiments. An
interesting line of future inquiry is joint optimal routing
by averaging over 104 random network realizations. Without and MAC layer scheduling/resource allocation for maximum
loss of generality, we assume that the network SNR is 20 dB spectral efficiency, while explicitly considering spatial reuse
in all experiments. Again, the running time of Algorithm B is and/or power control.
considerably (and consistently) lower than that of Algorithm
A. In particular, Algorithm B reduced the average time to R EFERENCES
obtain an exact optimal path by 89.08% – 94.26% relative to [1] I. F. Akyildiz, X. Wang, and W. Wang, “Wireless mesh networks: a
Algorithm A. survey,” Computer Networks, vol. 47, no. 4, pp. 445–487, Mar. 2005.
Finally, it is worth noting that the average running time [2] D. Chen, M. Haenggi, and J. N. Laneman, “Distributed spectrum-
efficient routing algorithms in wireless networks,” IEEE Trans. Wireless
of Algorithm B, in addition to the fact that it is entirely Commun., vol. 7, no. 12, pp. 5297–5305, Dec. 2008.
shortest-path-based, may suggest that the algorithm is suitable [3] D. S. J. De Couto, A. J. Bicket, and R. Morris, “A high-throughput path
for practical use. Furthermore, Algorithm B can be used as metric for multihop wireless routing,” in Proc. ACM MOBICOM, 2003.
[4] P. Djukic and S. Valaee, “Delay aware link scheduling for multi-hop
an efficient (polynomial-time) benchmark against which the TDMA wireless networks,” IEEE/ACM Trans. Networking, vol. 17, no.
performance of sub-optimal heuristics can be assessed, espe- 3, pp. 870–883, June 2009.
cially for a moderate-to-high number of nodes. We generate [5] R. Draves, J. Padhye, and B. Zill “Routing in multi-radio, multi-hop
wireless networks,” in Proc. ACM MOBICOM, 2004.
100 random 60-node network realizations, where the network [6] V. Genc, S. Murphy, Y. Yu, and J. Murphy, “IEEE 802.16j relay-based
SNR is set to 40 dB. For each realization, the optimal spectral wireless access networks: an overview,” IEEE Wireless Commun. Mag.,
efficiency is obtained using Algorithm B, and an approximate vol. 15, no. 5, pp. 56–63, Oct. 2008.
[7] P. Gupta and P. R. Kumar, “The capacity of wireless networks,” IEEE
solution is obtained using the distributed spectrum-efficient Trans. Inf. Theory, vol. 46, no. 2, pp. 388-404, Mar. 2000.
routing (DSER) algorithm introduced in [2]. DSER [2] oper- [8] J. L. Sobrinho, “Algebra and algorithms for QoS path computation and
ates by simply finding a shortest path from the source to the hop-by-hop routing in the Internet,” IEEE/ACM Trans. Networking, vol.
10, no. 4, pp. 541–550, Aug. 2002.
destination using (1+2𝛾 /𝑆𝑁 𝑅𝑙 ) as the link metric, where 𝛾 is [9] L. A. Wolsey, Integer Programming. John Wiley and Sons, Inc., 1998.
the path-loss exponent. Here, 𝛾 = 4. Fig. 4 depicts the relative [10] L. Xiao, M. Johansson, and S. P. Boyd, “Simultaneous routing and
gap to optimality introduced by DSER. On average, DSER resource allocation via dual decomposition,” IEEE Trans. Commun., vol.
52, no. 7, pp. 1136-1144, July 2004.
produced paths within 28% of the optimal spectral efficiency. [11] F. Xue, L.-L. Xie, and P. R. Kumar, “The transport capacity of wireless
In particular, the relative sub-optimality introduced by DSER networks over fading channels,” IEEE Trans. Inf. Theory, vol. 51, no.
ranged from 0 to 68.75%. Moreover, for 104 additional random 3, pp. 834-847, Mar. 2005.

Anda mungkin juga menyukai