Anda di halaman 1dari 6

Cross-layer Interference-aware Routing for Wireless Multi-hop Networks

Tamer ElBatt
Information and System Sciences Laboratory HRL Laboratories, LLC Malibu, CA 90265, USA

Timothy Andersen
Dept. of Mathematical Sciences Rensselaer Polytechnic Institute Troy, NY 12180, USA

ABSTRACT
In this paper we address the problem of interference-aware routing that tightly couples the design of the lower three layers of the ISO Open Systems Interconnection (OSI) protocol stack. This is primarily motivated by the observation that shortest path routing could potentially lead to degrading the single-hop throughput which constitutes an upper bound on the end-to-end multi-hop throughput. We introduce the concept of set-based routing in an attempt to incorporate interference into the routing decision as well as reduce the problem complexity. Towards this objective, we propose a novel algorithm that takes routing, scheduling and power control decisions for a set of interference-coupled transmitters. Furthermore, we discuss set coordination schemes for combating inter-set interference. Finally, we conduct a simulation study that shows considerable throughput improvement over a reference system that uses minimum hop routing and collision-free scheduling.

Categories and Subject Descriptors


C.2.1 [Computer-Communication Networks]: Network Architecture and DesignWireless Communication

General Terms
Algorithms, Performance

Keywords
Routing, scheduling, power control, interference, cross-layer design, simulation

1. INTRODUCTION
The problem of routing over multi-hop wireless networks has received considerable attention in the literature, primarily, along two thrusts: i) Efcient route discovery/maintenance under mobility conditions and time-varying topologies and ii) Developing link metrics to match a wide variety of performance objectives. Most of the protocols under the former thrust adopt the shortest path (SP) routing criteria widely employed in wireline networks [1]. Under the second thrust, a variety of routing metrics have been introduced for the purposes of energy-efciency, link stability, minimum delays, etc. Nevertheless, the problem of interference-aware routing

that tightly couples the MAC and routing decisions has not received adequate attention in the literature. [2] establishes bounds on the optimal throughput of interference-limited wireless multi-hop networks given the existence of an omniscient and omnipotent central entity. However, developing interference-aware routing protocols was left as an open problem. The problem of joint routing, scheduling and power control has been recently addressed in [3, 4, 5]. However, [3] assumes no interference between links in the same neighborhood sharing the same slot. On the other hand, [4] focuses on the simple setting of symmetric one-dimensional multi-hop networks. The work in [5] is the closest to ours, however, it hinges on the Gaussian approximation for interference [6]. This assumption facilitates separating the MAC and routing portions of the problem which considerably simplies the problem. In this paper, we present a solution to the problem without making any assumptions about the structure of interference known to be non-Gaussian in case of nite number of interferers. This work constitutes a step beyond our earlier work [7] to incorporate the impact of interference in the design of higher layer protocols. Our contribution in this paper is two-fold: i) Introduce a novel hop-by-hop set-based routing concept that incorporates interference into the routing decision of spatially close transmitters and ii) Introduce a joint routing, scheduling and power control (RSP) algorithm that solves the problem within each set. Motivated by the challenge of dening interference-aware link metrics, the complexity of the optimal RSP over the entire network [8] and the negligible interference among spatially far nodes, the solution proceeds through three steps: i) Construct interference-coupled transmitters sets, ii) Resolve intra-set interference via the RSP algorithm and iii) Resolve potential inter-set interference via set coordination. The paper is organized as follows: In section 2, the problem is motivated. In section 3, we highlight the challenges associated with interference-aware routing followed by a detailed description of set-based routing along with the RSP algorithm. The simulation results and discussion are given in section 4. Finally, conclusions are drawn in section 5.

2.

MOTIVATION

Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for prot or commercial advantage and that copies bear this notice and the full citation on the rst page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specic permission and/or a fee. IWCMC06, July 36, 2006,Vancouver,British Columbia,Canada. Copyright 2006 ACM 1-59593-306-9/06/0007 ...$5.00.

In this section, we show that handling MAC and routing decisions in isolation could lead to throughput degradation in wireless multi-hop networks. This is conrmed with the aid of a simple example. Consider a wireless network consisting of 13 stationary nodes with connectivity shown in Figure 1. For this example, we assume time is slotted and the channels are constant over a time slot. Moreover, we assume two source-destination (S-D) pairs (S1 , D1 ) and (S2 , D2 ) with identical trafc demands and there is always a packet in the queue of each source ready for transmission. We compare the average slot throughput and end-to-end (E2E) throughput of three routing policies. The slot throughput is

dened as the average number of successful transmissions per slot. In this section, the E2E throughput is dened as the average number of packets that are successfully transferred from each source to its respective destination over d slots.
BAAB PIIP
D1 D2

3. INTERFERENCE-AWARE ROUTING 3.1 Challenge


Our focus is to incorporate interference awareness mechanisms on top of table-based routing. Thus, we assume that next hop nodes residing on all possible routes to destination (not only MH routes) are known at each hop. Two nodes are single-hop neighbors if they can communicate using maximum power assuming no interference. Incorporating interference into on-demand routing decisions is out of the scope of this paper. We consider n nodes which communicate only via the wireless medium and are indexed with unique IDs from 1 to n. All nodes share the same frequency band and time is divided into slots that are grouped into frames. Each frame is of xed length and divided into a control sub-frame and a data sub-frame which consists of d slots. We focus on unicast trafc. Next, we identify the fundamental challenge associated with dening link-based interference-aware routing metrics. Under SP routing, the path length (which depends on the link metric) is the only factor that decides the best route between any S-D pair. Various examples of link metrics in the literature, namely Euclidean distance, residual battery charge, and buffer occupancy, depend solely on the two nodes forming the link. They are independent of the existence of other S-D pairs or their SP routes. This, in turn, has led to the notion of link metrics and link-based routing. However, interference depends on the existence of other sources/intermediate relays and their spatial separation. Hence, the routing decision of a given S-D pair becomes coupled to the routing decision of other SD pairs. Accordingly, the notion of a link metric that incorporates interference becomes impossible as illustrated by the following example: Assume node a is transmitting to next hop b and node u is transmitting to next hop v as shown in Figure 2(a). According to the non-linear decay of power with distance, governed by Pr (z) = Pt z where Pt is the transmitted power, z is the distance between transmitter and receiver and is the path loss exponent, the amount of interference at node v from transmitters u other than u is given by: Iv = Pab zav . If node a was transmitting to a different next hop (say c) as shown in Figure 2(b), then the amount of interference seen at node v would be different: u Iv = Pac zav . Thus, the interference introduced to link uv
b v

I1 I3 S1

I2

I1

I5

I3 S1 D2

I5

(a) Schedule: S 1, S 2 , I 0 ,I

S2
0

S2

(b) Schedule: S 1, S 2, I 1 I 2

I4 I1

I5 S1

S2

(c) Schedule: S 1 S 2, I 3I 5, I 4 I 6 S 1 S 2

Figure 1: Schedules associated with 3 routing policies. The notation used in (b) for the transmission schedule S1 , S2 , I1 I2 can be interpreted as: node S1 transmits in slot 1, node S2 transmits in slot 2 and nodes I1 and I2 transmit simultaneously in slot 3. First, we consider minimum hop (MH) routing which is a special case of SP routing when all links have unit metrics. Based on the topology, there are two MH routes for each pair and we consider rst the MH routes that pass through node I0 as shown in Figure 1(a). It is evident that I0 constitutes an interference bottleneck at the MAC layer. This is a direct consequence of ignoring the interaction among the chosen paths through the amount of interference a certain path may introduce to others. We refer to this problem as interference-induced congestion which constitutes a fundamental challenge unique to wireless networks. Thus, the two source nodes can not simultaneously transmit to I0 in the same slot, and neither one can transmit while I0 is transmitting. Moreover, I0 can forward only one packet to either destination at a time. This yields an average slot throughput of 1 transmission per slot. Moreover, this policy consumes 4 slots for a single packet transfer from each source to its respective destination (referred to as the Transmission Cycle). Hence, we conclude that the above routing policy can transfer d packets, from each source to its destination, over d slots. 4 Second, we consider another MH routing policy where node S 1 forwards to I1 and S2 forwards to I2 as shown in Figure 1(b). We assume that nodes I1 and I2 are spatially close enough to: i) prevent simultaneous reception from the source nodes and ii) prevent I 1 from receiving while I2 is transmitting and vice versa. In addition, we assume that nodes D1 and D2 are spatially far enough to allow simultaneous reception of their packets. Clearly, this policy should yield better performance due to choosing different next hops for S 1 and S2 . However, the next hops are very close which still restricts spatial reuse. The transmission cycle becomes 3 slots, the average 4 slot throughput 3 transmissions per slot and the E2E throughput d increases to 3 packets, i.e. 33% improvement over Figure 1(a). Finally, we examine the performance of a routing policy that uses longer paths, yet, spatially far enough to allow simultaneous transmissions over both paths. Assume that S1 follows the 3-hop path through I3 and I4 , whereas S2 follows the path through I5 and I6 as shown in Figure 1(c). If slot reuse is allowed on the same path, then the average slot throughput can be shown as 3 transmissions per slot and the transmission cycle becomes 2 slots. Moreover, the E2E throughput turns out to be d packets, i.e. two-fold improve2 ment over the rst policy and 50% better than the best MH routing policy in Figure 1(b). This conrms the fundamental role interference plays in limiting single-hop throughput and, consequently, E2E throughput in wireless multi-hop networks.

Iv

c a (a)

c u a (b)

Figure 2: The challenge of dening an interference-aware link metric: interference introduced to link uv depends on the routing decision of node a which, in turn, depends on the interference caused by node u (needed to compute its link metric) depends on the routing decision of transmitter a which, in turn, depends on the routing decision of transmitter u. This suggests that there is a fundamental hurdle towards dening link metrics that explicitly account for interference. In addition, it constitutes a key observation that motivates the notion of set-based routing presented next where the MAC and routing decisions are taken jointly for a set of spatially close transmitters (as opposed to link-based routing where the routing decision is taken for each transmitter independently).

3.2

Cross-layer Set-based Routing

In this section, we incorporate interference into hop-by-hop routing decisions through the concept of set-based routing where deciding next hops is taken jointly for a set of transmitters. The proposed

y y yyyy y y yyy yy y yyyy y yy yy yyy yy y yyyy y y yy yyy yy

WWXX

I3

`YY`

rqqr

I0

I2

UVVU

SSTT

h g gghh gh

D1

I6

7887

xwwx

GHHG

EFFE

6556

CDDC

4334

(''(

piip

RQQR

tsst



baab

uuvv

cddc

efeef



I0

)00)

I6

I0

I2

9@@9

! !"" !!

%&&% 



I4

I4

2112



D1

D2 I6

#$$# 

Iv

algorithm is executed on a frame-by-frame basis where a packet can progress at most one hop in each frame. Thus, given a set of K S-D pairs at the beginning of a frame (where source nodes may be original sources of packets or intermediate relays), we rst investigate the problem of set construction and its associated trade-offs in light of the clustering algorithms introduced in the literature. Second, we introduce the joint routing, scheduling and power control (RSP) algorithm which constitutes an important contribution of the paper. Finally, we discuss candidate set coordination schemes to combat inter-set interference.

3.2.2

Joint Routing-Scheduling-Power Control

3.2.1

Set Construction

The problem of grouping transmitters depending on their interference coupling exhibits similarity to the problem of clustering in mobile ad hoc and sensor networks [9]. In contrast to the known motivation for clustering (e.g. local processing and communications for scalability and reduced overhead), it is solely motivated here by interference. For instance, [5] groups transmitters in a prespecied geographical region in the same cluster. However, intercluster interference was approximated as static ambient noise. In this paper, we adopt a simple topology-based criteria for dening interference-coupled transmitters (ICT) sets, such that transmitters within H hops from a specic transmitter belong to the same ICT set. Otherwise, they belong to different sets. Recall that transmitters (original sources/intermediate relays) and nal destinations are the only information available at this point of the algorithm, i.e. next-hop receivers are not known yet, and hence, set construction is solely based on transmitters spatial separation. Despite its simplicity, the topology-based criteria could dynamically control the set size, depending on the number of transmitters and their spatial separation, via adapting the parameter H. This gives rise to an interference model that explicitly accounts for interference caused by transmitters who are within the same set. Otherwise, interference is handled indirectly as discussed in section 3.2.3. This model is considered more realistic than the circular interference range typically used in 802.11 modelling. One way to construct ICT sets according to a pre-specied parameter H is described next. Given a set of K S-D pairs at the beginning of a frame, the algorithm goes through all nodes, in an ascending order of their IDs, where it skips nodes who have nothing to transmit in the current frame. On the other hand, nodes who wish to transmit in the current frame disseminate their IDs to their H-hop neighbors via H-hop ooding. By the completion of the algorithm, each node in the network would have a list of up to H-hop away transmitters. Nodes who have nothing to transmit simply ignore their respective lists whereas transmitter k considers its H-hop transmitters list as ICTk set, becomes the leader of this set and noties set members. Thus, the above algorithm constructs K ICT sets, one per transmitter. This, in turn, leads to overlapping among most of the constructed sets (i.e. transmitters who belong to multiple ICT sets at the same time). Coordination schemes for handling set overlapping are discussed in section 3.2.3. Notice that transmissions in the ICT set construction phase are carried over the control sub-frame at the beginning of each frame, possibly using TDMA contention-free scheduling. This is justied by the fact that control messages are considerably smaller than data packets and, hence, contention-free transmission should not constitute signicant waste of resources. Finally, it can be shown that the number of control messages per frame associated with ICT set construction scales linearly with the number of transmitters K H 1 and exponentially with H, as O(K N 1 ), where N is the average N number of single-hop neighbors.

In this section, we introduce a sub-optimal RSP algorithm that solves the problem within a given ICT set. The objective is threefold: rst, to determine the next hop subject to the following routing constraints: (i) The next hop of any transmitter should be among its neighbors and (ii) Length of the path from Sk to Dk should be less than a pre-specied threshold, k , which relaxes the MH criteria widely used in the MANET literature. Second, to decide which link is activated in which slot subject to the scheduling constraints: (iii) No simultaneous transmission/reception by a node in any slot (self-interference problem) and (iv) No multiple simultaneous transmissions to the same receiver in any slot (commonreceiver problem). Third, to specify the set of powers needed to satisfy the power control problem constraints: (v) The signal-tointerference-and-noise-ratio (SINR) is greater than a pre-specied threshold (power admissibility condition) and (vi) The peak power constraint per node (Pmax ). Next, we highlight the main features of the proposed algorithm followed by a detailed description of the interaction of various parts of the algorithm. The essence of RSP is to construct links and schedules that pack as many transmissions as possible in each slot while guarantee convergence of the continuous, iterative, distributed, power control (DPC) algorithm, introduced in [10], to the minimum power vector in all slots: Ptj (T + 1) = min[Pmax , Pt (T )] SIN Rj (T ) j (1)

where Ptj is the power transmitted from node j to its next hop neighbor, SIN Rj is measured at the receiver of node j, is a minimum requirement on the SINR for successful reception and T is the iteration number. Thus, the routing and scheduling portions of the algorithm are search-based. This is attributed to their combinatorial nature [8] and lack of a tractable mathematical structure that lends itself to analytical optimization techniques. On the other hand, DPC is an iterative algorithm that is known to converge exponentially fast to the minimum power vector, if one exists. Hence, techniques for speeding up the routing and scheduling search process are sought. Towards this objective, we incorporate the following features. First, the routing search process commences with the MH policy since minimizing the path length is a desirable criteria in the solution. If there are multiple MH paths, we resolve the tie via random selection. The natural question that arises next is: In what order should the routing policies be examined? The signicance of this question stems from the fact that neighbors of a source/intermediate relay differ in the length of the path on which they reside and the amount of interference they get exposed to. Thus, we argue that neighbors who satisfy the path length constraint (ii), along with the loop-freedom constraint, should be ordered according to their path lengths to destination, such that neighbor(s) on MH path(s) are examined rst and neighbors on longer paths are examined later. This ranking is instrumental in guiding the search process throughout the routing policy space, and, hence, nding good sub-optimal solutions in a timely manner. Second, the scheduling search process commences with the All Transmissions in a Single Slot (ATSS) policy followed by policies that tend to distribute transmissions evenly among slots in the frame. The rationale behind this choice is to pack as many transmissions as possible in a slot in order to maximize the single-hop MAC throughput which constitutes an upper bound on the multi-hop throughput. If this leads to empty slots in the frame, then the generated schedule, or part of it, could be repeated using next packets in the queues ready for transmission. Figure 3 shows a owchart that demonstrates the interaction of the routing, scheduling and power control portions of the algorithm.

As indicated earlier, the algorithm starts with examining MH routing and ATSS scheduling as directed by the set leader. Accordingly, all transmitters in the set send control information to the set leader, in a contention-free manner, that denes the set of links. This is essential to solve the self-interference and common receiver problems. Accordingly, the set leader examines constraints (iii) and (iv) (i.e. scheduling constraints) for all slots in the frame. If one or both constraints are violated, the scheduling algorithm defers conicting transmissions to a future slot, using the heuristic in [7] which examines the two constraints in sequence and defers users transmissions to resolve conicts. Otherwise, the algorithm proceeds to the power control portion with a set of single-hop links along with their slot assignments. The DPC algorithm, executed by transmitter-receiver pairs in each slot, examines their power admissibility judged by the convergence of (1). Notice that the operation

tion, all routing policies that satisfy constraints (i) and (ii) and are loop-free are examined until either a solution is found or another scheduling policy is examined. This choice is based on the premise that routing could reduce interference, like scheduling, yet without sacricing the MAC throughput, which constitutes an upper bound on the E2E throughput. Notice that a new scheduling policy is generated from an existing one via deferring the link(s) with minimum SINR [7] in the last DPC iteration of violating slot(s) in an attempt to lower the level of interference. This heuristic enables the remaining links to converge to the minimum power vector quite fast.

3.2.3

Set Coordination

Given ICT set at the beginning of frame i

Initial Routing and Sched. Policies Routing: MH Scheduling: ATSS

Are Conflicting transmissions are deferred to next slot(s) NO constraints (iii) and (iv) satisifed in all slots ?

YES Pick another Routing Policy in increasing order of path length s.t. constraints (i) & (ii) NO Are all slots power admissible? NO Are all routing policies examined for the same scheduling policy? YES Source nodes exit with: - Routing Policy - Link Scheduling Policy - Transmission Powers YES Pick another Scheduling Policy via deferring the transmisisons with min. SINR in violating slots

It is evident from section 3.2.1 that ICT sets could be overlapping, i.e. one or more transmitters may belong to multiple ICTs. Set overlapping gives rise to the so-called inter-set interference which leads to coupling different ICT sets. Executing the RSP algorithm for these sets simultaneously is problematic since transmitter(s) in the overlapped regions would receive conicting orders from respective set leaders. This is equivalent to merging the sets, yet, the leaders of the original sets are trying to nd solutions for their respective sets without any coordination. One way to circumvent this hurdle is to enable the merged sets to somehow detect the overlapping and then elect a new leader for the larger set who runs a single RSP algorithm. However, this is achieved at the expense of two drawbacks: rst, executing RSP for larger sets implies higher computational overhead due to larger routing and scheduling policy spaces. Moreover, it could lead to merging a chain of ICT sets to a single set that covers the entire network. Second, it contradicts one of the fundamental premises in the paper that spatially far transmitters, whose mutual interference is negligible, should belong to different sets. In this section, we embrace an alternative approach that avoids set merging via the notion of set scheduling (or coordination) that is carried out in the control sub-frame. Thus, the problem boils down to coordinating the execution of the RSP algorithm for different ICT sets such that: i) It is executed for overlapping ICT sets at different times and ii) Executed simultaneously for non-overlapping ICT sets. The problem of determining the minimum length set schedule subject to the above two constraints is a combinatorial problem that requires global information about set overlapping at a central controller. It can be mapped to a graph coloring problem via constructing a conict graph for ICT sets. We limit our attention here to a sequential coordination scheme that satises the rst constraint only. Adopting such coordination scheme constitutes a simple rst step towards the problem and is justied by our focus on the proposed cross-layer framework and RSP algorithm as the main contributions. Simulation results presented in the next section conrm the profound impact that setbased routing (RSP+sequential coordination) has on the network throughput. According to the proposed coordination scheme, the RSP algorithm is executed for all ICT sets sequentially, in an ascending order of leader IDs, irrespective of their overlap. Distributed operation is achieved via exchanging leader IDs throughout the network with the aid of efcient information dissemination schemes, e.g. probabilistic ooding. Notice that a transmitter that belongs to multiple ICT sets determines it routing, scheduling and power control decisions through the RSP execution of the set reaching its turn rst. Afterwards, its solution is handled as a constraint during the RSP execution of the other sets. This represents a key feature of this scheme which resolves inter-set interference without the need to merge overlapped sets. Finally, it can be shown that the control overhead of sequential coordination grows exponentially with H since each set leader is responsible for ooding its

Execute the DPC algorithm for each slot in frame i

Go to next frame (i = i + 1)

Figure 3: Flowchart of the Intra-set RSP Algorithm of DPC requires receivers to send their SINR measurements back to their respective transmitters at the end of each iteration. This is achieved using a separate feedback channel that can be used in a contention-free manner since the feedback messages are considerably smaller than data packets. If all slots turn out to be power admissible, then the set leader announces the routing, scheduling and power control solutions. Otherwise, another routing and/or scheduling policy is examined for power admissibility. This gives rise to an important question: Which of the two policies, routing or scheduling, should be given precedence in the search process? Motivated by the observation that routing can circumvent the negative impact of interference without degrading the MAC throughput, we give precedence to examining routing policies over scheduling policies. More specically, for a scheduling policy under examina-

set solution over H hops in order for the leaders of overlapped sets to utilize such information in solving their respective sets.

4. PERFORMANCE EVALUATION 4.1 Simulation Setup


In this paper, we conduct simulation experiments using the ns-2 simulator. We consider a uniform network topology where n = 64 xed nodes are placed across a square area of length 1000 meters. The square is split into 64 smaller squares where the location of each node is selected randomly within each of these squares. Each node is supported by an omni-directional antenna that radiates energy according to an isotropic pattern and the path loss exponent = 4. The number of data slots per frame is set to d = 5 where each slot is of duration 6 msec. We primarily focus on xed multihop wireless networks (e.g. mesh networks) and, hence, mobility is not modelled. However, this assumption can be relaxed to the class of low mobility MANETs where the link gain matrix changes over a time scale much larger than the time scale over which the proposed algorithm operates, namely a frame. A source node generates packets of length 512 bytes according to a Poisson process with rate packets/sec. We assume that each node has a queue of arbitrarily large size since our objective here is to capture packet losses attributed to interference. The maximum radio transmit power is set to 20 dBm which translates to a range of approximately 300m in the absence of interference. The SINR threshold for successful reception at a receiver () is set to 5 dB. The receiver thermal noise power is assumed to be -90 dBm. The maximum number of iterations in the DPC algorithm is set to 20. We assume three S-D pairs where source nodes are separated from their respective destinations by approximately 1000 meters on the average in order to emphasize the role of interference contributed by intermediate nodes on multi-hop paths. Motivated by the computational complexity of solving DPC for large number of links along with the exponential growth of control overhead per frame with H, we limit H to small values (H = 2) in this set of experiments. The network load is varied via increasing the packet arrival rate per source node () from 10 packets/sec to 360 packets/sec. Each simulation run is carried out for the duration of 900 seconds.

4.2 Reference System


In this section, we describe the reference system (REF) used as a benchmark to gauge the performance gains achieved by jointly designing the MAC and routing algorithms. First, we assume the transmission power is xed at Pmax and there is no provision for dynamic adaptation. Second, we assume that the routing and multiple access decisions are taken in isolation. A table-based routing protocol is executed by each source/intermediate relay using the classical link-based MH metric. Ties between multiple MH paths are resolved via random selections. Once the routing decision is made for all transmitters at the beginning of each frame, then it is the responsibility of the scheduling algorithm to resolve the contention among the constructed links. Hence, we adopt maximal slot scheduling proposed in [11] which attempts to create collision-free schedules via satisfying the following constraints: i) A node is not allowed to simultaneously transmit and receive in the same slot, ii) A node is not allowed to receive from more than one transmitter in the same slot and iii) A receiver should not be a neighbor of any other transmitter.

4.3 Performance Results


The prime focus of this paper is to assess the throughput improvement achieved by the joint design of MAC and routing over the reference system. Comparing the control overhead associated with the two systems requires a detailed simulation of the control

sub-frame which is out of the scope of this paper. The performance metric used to compare the two systems is the E2E throughput dened as the long-run average number of data packets that reach their respective nal destinations successfully per second. In order to better understand the trade-offs involved, we examine three routing policies. These policies are generated via widening the scope of the routing policy space (through varying the path length constraint k ). Recall that the routing search process in the RSP algorithm ranks MH paths in the highest rank and the rank decreases as the path length increases. Accordingly, the rst policy, referred to as Highest Rank Paths Only (HRPO), limits search in the routing policy space to neighbors residing on MH paths. This policy may show performance gains only if there are multiple MH paths between some S-D pairs, which is typical in many networks with moderate connectivity. The second policy, referred to as Second Rank Paths Also (SRPA), widens the search scope to accommodate MH paths in the highest rank along with longer path(s) in the second rank. This policy trades path length for MAC throughput in an attempt to improve the E2E throughput. Finally, the third policy, referred to as All Rank Paths (ARP), is an extreme one that considers neighbors on all possible paths between the source and destination as potential next hops, irrespective of their associated path lengths. Accordingly, MH routing and ARP policies can be viewed as the extremes. At one end of the spectrum, MH routing attempts to minimize path lengths irrespective of the MAC throughput. At the other end of the spectrum, the ARP policy attempts to maximize the MAC throughput irrespective of path lengths. Simulation results conrm that neither extreme constitutes a favorable design choice. First, we compare the long-run average number of successful single-hop transmissions per slot under the four schemes. The importance of this experiment stems from the fact that this parameter reects the MAC Throughput. It can be easily noticed from Figure 4 that the reference system yields the lowest MAC throughput due to ignoring the negative impact MH routing may have on the MAC performance. On the contrary, the HRPO and SRPA policies show considerably higher MAC throughput. Notice that HRPO improves the MAC throughput over REF by a factor of 25% at heavy loads. It is crucial to notice that this is achieved while preserving the MH routing criteria since the HRPO policy attempts to exploit the spatial separation of next hop nodes to pack as may transmissions as possible in each slot. Moreover, it can be noticed that as longer paths are considered, the SRPA policy improves the MAC throughput over REF by a factor of 38% at heavy loads. This is attributed to enlarging the routing search space which creates more room for spatially separating the paths of various S-D pairs. Under the ARP policy, the MAC throughput increases at low to moderate loads and then starts to decrease under high loads. We attribute this phenomena to routing based solely on MAC throughput which could lead to forwarding packets to directions far from destination. This may lead to following arbitrarily long paths, as shown later, which imposes higher load on the network and starts to have negative impact on the MAC throughput around 150 packets/sec/source node. Next, we compare the long-run average path length under the four policies as shown in Figure 5. This measure reects the price paid for improving the MAC throughput. It is straightforward to notice that the reference system yields the lowest average path length due to adopting the MH routing criteria. Moreover, policies that improve the MAC throughput, namely HRPO and SRPA, have average path lengths similar to or longer than the reference system. Finally, the ARP policy has the longest path length on the average as expected. Figures 4 and 5 conrm that the interplay between MAC throughput and path length is what determines the net E2E throughput.

3 REF HRPO SRPA ARP

2.8

2.6

2.4

2.2

1.8

1.6

1.4

1.2

50

100 150 200 250 300 Packet Arrival Rate per Source Node (packets/sec)

350

400

Figure 4: Average Number of Successful Transmissions per Slot for four Routing Policies Figure 6 shows the E2E throughput performance under the four policies. First, we notice that the reference system yields low performance due to ignoring the trade-off between MAC throughput and path length. Second, the HRPO and SRPA policies give the highest E2E throughput due to improving the MAC throughput subject to a constraint on the path length. Third, the ARP policy yields the worst throughput performance due to following spatially far, yet excessively long, paths from source to destination. Finally, we notice that the HRPO policy outperforms REF by a factor of 50% at heavy loads and decreases to 35% at moderate loads. On the other hand, the SRPA outperforms the reference system by a factor of 34% for moderate and heavy loads. Notice that although the SRPA policy is inferior to HRPO, it still outperforms the reference system. Referring to the relative performance of HRPO, SRPA and ARP we observe that there is a turning point in the behavior of the RSP algorithm, that is directly related to the path length constraint and worth further investigation. Under HRPO and SRPA, where routing is restricted to MH and slightly longer paths, there could be room for overall performance improvement as demonstrated. On the other hand, as we approach the ARP policy (via relaxing the path length constraint), improving the MAC throughput becomes out weighed by the lengthy paths followed from source to destination. Thus, we conclude that by appropriately limiting the scope of the routing policy space, the proposed cross-layer framework could achieve signicant performance improvement over MH routing.

Proc. ACM MOBICOM, Sept. 2003. [3] R. Bhatia et al. On Power Efcient Communication over Multi-hop Wireless Networks: Joint, Routing, Scheduling and Power Control, IEEE INFOCOM, March 2004. [4] B. Radunovic et al. Joint Scheduling, Power Control and Routing in Symmetric, One-dimensional Multi-hop Wireless Networks, WiOpt Workshop, March 2003. [5] R. Cruz and A. Santhanam Optimal Routing, Link Scheduling and Power Control in Multi-hop Wireless Networks, Proc. IEEE INFOCOM, April 2003. [6] T. Rappaport, Wireless Communications Principles and Practice. New Jersy:Prentice-Hall Inc., 1996 (2nd Ed. 2002). [7] T. ElBatt and A. Ephremides Joint Scheduling and Power Control for Wireless Ad hoc Networks, IEEE Transactions on Wireless Communications, vol. 3,no. 1, Jan. 2004. [8] T. ElBatt and T. Andersen A Cross-layer Framework for Multiple Access and Routing Design in Wireless Multi-hop Networks, under submission, Dec. 2005. [9] D. Baker and A. Ephremides The Architectural Organization of a Mobile Radio Network via a Distributed Algorithm, IEEE Transactions on Communications, vol. 29, No. 11, Nov. 1981. [10] G. Foschini and Z. Miljanic A Simple Distributed Autonomous Power Control Algorithm and its Convergence, IEEE Transactions on Vehicular Technology, vol. 42, No. 4, Nov. 1993. [11] I. Cidon and M. Sidi Distributed Assignment Algorithms for Multihop Packet Radio Networks, IEEE Transactions on Computers, vol. 38, No. 10, pp. 1353-1361, Oct. 1989.
20 REF HRPO SRPA ARP

Average Number of Successful Transmissions per Slot

18

Average Number of Hops from Source to Destination

16

14

12

10

5. CONCLUSIONS
In this paper we introduced a cross-layer multiple access and routing algorithm for interference-limited wireless multi-hop networks. Our main contribution is to incorporate interference into the routing decision via the set-based routing framework that solves the problem for spatially close transmitters as opposed to the classical link-based routing paradigm. Moreover, we introduced a novel joint routing, scheduling and power control (RSP) algorithm that handles intra-set interference. We conducted a simulation study that compares the cross-layer MAC and routing framework to a reference system where MAC and routing decisions are taken in isolation. Results exhibit performance improvement up to 50% for a variety of routing policies. Finally, this work opens room for exploring interference-aware on-demand routing protocols and developing distributed minimum-time set coordination schemes.

50

100 150 200 250 300 Packet Arrival Rate per Source Node (packets/sec)

350

400

Figure 5: Average Path Length for four Routing Policies

45 REF HRPO SRPA ARP

40

Average EndtoEnd Throughput (packets/sec)

35

30

25

20

15

10

6. REFERENCES
[1] D. Bertsekas and R. Gallager, Data Networks. New Jersy:Prentice-Hall Inc., 1987 (2nd Ed. 1992). [2] K. Jain, J. Padhye, V. Padmanabhan and L. Qiu, Impact of Interference on Multi-hop Wireless Network Performance,

50

100 150 200 250 300 Packet Arrival Rate per Source Node (packets/sec)

350

400

Figure 6: Average E2E Throughput for four Routing Policies

Anda mungkin juga menyukai