handling asymmetry in power heterogeneous ad hoc networkskrish/comnet-power.pdf · heterogeneous...

Computer Networks 51 (2007) 2594–2615

www.elsevier.com/locate/comnet

Handling asymmetry in power heterogeneous ad hoc networks

Vasudev Shah, Ece Gelal *, Srikanth V. Krishnamurthy

Department of Computer Science and Engineering, University of California, Riverside, CA 92521, USA

Received 4 May 2006; received in revised form 19 September 2006; accepted 20 November 2006Available online 20 December 2006

Responsible Editor: M. Venkatachalam

Abstract

Traditional MAC and routing protocols, which are primarily designed for homogeneous networks wherein all nodes trans-mit with the same power, suffer performance degradations when employed in power heterogeneous networks. The observeddegradations are due to link asymmetry, which arises as high power nodes that do not sense the transmissions of low powernodes can potentially initiate transmissions that interfere with the low power communications. Link layer asymmetry in powerheterogeneous networks not only disrupts the functioning of the routing protocol in use, but also results in unfairness in med-ium access. In this paper, we develop a cross-layer framework to effectively address the link asymmetry problem at both theMAC and the routing layers. At the MAC layer, the framework intelligently propagates low power control messages to thehigher power nodes, so as to preclude them from initiating transmissions while there are low power communications in pro-gress within their sensing range. At the routing layer, the framework facilitates the efficient use of unidirectional links. Weperform extensive simulations to study the performance of our proposed framework in various settings, and show that theoverall throughput in power heterogeneous networks is enhanced by as much as 25% over traditional layered approaches.In addition, we show that our schemes are also beneficial in power homogeneous settings, as they reduce the extent of falselink failures that arise when the IEEE 802.11 MAC protocol is used. In summary, our framework offers a simple yet effectiveand viable approach for medium access control and for supporting routing in power heterogeneous ad hoc networks.� 2006 Elsevier B.V. All rights reserved.

Keywords: Power control; Power-heterogeneous ad hoc networks; Link asymmetry; Cross-layer design; MAC; Routing

1. Introduction

Emerging ad hoc networks are likely to consist ofdevices with varying capabilities. One could envi-sion low power sensor nodes and wireless hand-helddevices integrated into a single network with higherpower wireless devices such as laptops, wireless rou-

1389-1286/$ - see front matter � 2006 Elsevier B.V. All rights reserved

doi:10.1016/j.comnet.2006.11.016

* Corresponding author. Tel.: +1 909 3798258.E-mail address: [email protected] (E. Gelal).

ters or wireless devices mounted on vehicles, pow-ered by inside alternators. In such heterogeneousnetworks nodes are likely to transmit at differentpower levels, thereby causing communication linksof varying range. In such networks link asymmetry

is likely to be the norm. Link asymmetry may alsoappear in ad hoc networks when power control isemployed, in order to reduce energy consumptionor to enhance spatial re-use in the network (as in[11,20]). With link asymmetry, the transmission of

.

mailto:[email protected]

V. Shah et al. / Computer Networks 51 (2007) 2594–2615 2595

a lower power node might not be received (orsensed) at a higher power node, while communica-tion in the reverse direction could be feasible. As aresult, traditional MAC and routing protocols thatimplicitly assume that links are bi-directional willeither fail or perform poorly. At the MAC layerthe hidden terminal problem is exacerbated [26];routing becomes challenging due to presence of uni-directional links [19,24,3].

There have been a plurality of research efforts onalleviating the effects of link asymmetry in ad hocnetworks. It has been shown in [26] that the perfor-mance of the IEEE 802.11 MAC protocol degradesin the presence of link asymmetry, since low powernodes cannot acquire the channel for sufficientdurations. Link asymmetry at the MAC layer hasalso been studied in [4,18]; however, these studiesare limited to certain scenarios with strict assump-tions. In parallel, [19,24,3,2,31] propose methodsfor routing in the presence of unidirectional links;however, most of these studies ignore interactionswith the MAC layer. Specifically, they either (i)employ the traditional IEEE 802.11 MAC protocol,which is inefficient with link asymmetry, or (ii) theyassume an ideal MAC protocol which can onlyfunction effectively in the presence of unidirectionallinks. Finally, there are proposals to build link layertunnels in order to hide unidirectional links fromthe higher layers [9,12]; again, MAC layer effectscan impact these solutions.

The prior work in [26] considers the propagationof MAC layer control messages (in particular theIEEE 802.11 MAC protocol is studied) in order toalleviate the effects of asymmetry. However, thework considered variants of flooding for the abovepropagation and the overhead incurred was shownto be prohibitive; as a result, performance enhance-ments were not achieved. We first examine if reduc-ing this overhead via some simple strategies canimprove the MAC layer performance. In particular,we study the effect of (i) using an intelligent broad-casting scheme to quell unnecessary broadcasts,and, (ii) reserving the bandwidth for multiple datapackets with a single RTS/CTS exchange (themulti-reservation scheme). Our simulations showthat these extensions offer fairly limited perfor-mance enhancements (as compared to the legacyIEEE 802.11 protocol) in terms of the MAC layerthroughput.

The above studies lead us to believe that a furtherreduction in the MAC layer control message propa-gation overhead could yield higher performance

dividends. In order to do so, as opposed to relyingon broadcasting strategies like before, we design across-layer solution, wherein, MAC layer controlmessages that are transmitted with low powers arerouted to beyond the one-hop neighborhood of thelow power transmitter. As before, the goal is toinform and silence the high power nodes that, beingoblivious of low power communications within theirsensing range, can potentially initiate transmissionsthat will collide with existing low power communica-tions. However, with this approach the functionalityis achieved with much lower overhead. In addition,our framework implicitly facilitates the tunneling ofMAC and routing control packets via a reverse path

(also referred to as the inclusive cycle [3]) that spans aunidirectional link. Tunneling supports the effectiveutilization of the unidirectional links at the rout-ing layer. In summary, our framework effectively:

• Eliminates MAC layer inefficiencies in power het-erogeneous ad hoc networks, thereby increasingthroughput achieved at the MAC layer to a valuethat commensurates with that of a power homo-geneous network.

• Identifies and effectively utilizes unidirectionallinks at the routing layer, thereby improving theperformance in terms of throughput perceivedat the higher layers (as compared to that with tra-ditional routing protocols under power heteroge-neous conditions).

We implement our framework in two steps. First,we modify the MAC layer to incorporate basic rout-ing functionalities; this results in our topology-aware

CTS propagation (TACP) scheme. With TACP,CTS messages are ‘‘routed’’ to the higher powernodes prior to any data exchange involving lowpower nodes. We combine TACP with the previ-ously mentioned multi-reservation scheme. Thesetwo schemes are independent, and provide comple-mentary improvements in performance. Second, weextend TACP to tunnel MAC and routing controlpackets in the reverse direction of a unidirectionallink on a path that spans this link. (The structurethat is formed for propagating the CTS message isused for this purpose.)

We perform extensive performance evaluations ofour framework via simulations. First, we eliminatehigher layer artifacts and examine the performanceimprovements exclusively at the MAC layer. In thisstage, we do not tunnel packets or perform routing.We observe that the MAC throughput of the low

2596 V. Shah et al. / Computer Networks 51 (2007) 2594–2615

power nodes improves by approximately 24% ascompared to the IEEE 802.11 MAC protocol. Next,we study the benefits offered by our framework asperceived at the transport layer (UDP). We showthat the number of false link failures reported tothe routing layer is reduced by up to 20%. In addi-tion, the overall packet delivery ratio observed atthe transport layer improves by up to 25% as com-pared with the traditional layered structure (withthe IEEE 802.11 MAC protocol and AODV in place).

The rest of this paper is organized as follows: InSection 2, we describe related previous work. In Sec-tion 3, we briefly revisit the problems due to linkasymmetry. In Section 4, we describe our methodol-ogies in detail; in particular, we discuss: (a) specificMAC layer enhancements that were initially consid-ered, and (b) our cross-layer framework. In Section5, we present our simulation results and discuss theachieved performance enhancements. Our conclu-sions form Section 6.

2. Related previous work

In this section, we describe related work thataddress the effects of link asymmetry at variouslayers. Unlike most of these previous efforts, weconsider a truly multifarious network wherein thenodes could have different maximum transmissionpower capabilities. Furthermore, our cross-layerframework enhances performance at both MACand routing layers unlike almost all previousschemes which, in their design, ignore the effectsof link asymmetry at other layers.

2.1. Link asymmetry due to power control at the

MAC layer

Power control in MANETs has been explored in[5,11,20,26]. The authors in [11,20] propose power-controlled MAC protocols that incorporate the col-lision avoidance mechanism of the legacy IEEE802.11 MAC protocol. With both protocols, therequest-to-send (RTS) and the clear-to-send (CTS)frames are transmitted with a maximum presetpower level, so that all nodes within the maximumrange can hear the transmissions. Data frames arethen communicated using the lowest power level thatsuffices for the intended communication to succeed.Both schemes avoid the effects of asymmetry byemploying maximum-power transmission of controlframes. These protocols are not able to exploit thespatial re-use gains that are potentially possible with

power control. Furthermore, they are not applicablein networks consisting of multifarious nodes withdifferent maximum transmission power capabilities.

A busy tone-based power control protocol is pro-posed in [33], where the busy tone is transmitted ona separate control channel at the maximum powerlevel. Each neighbor estimates the channel gainfrom the busy tone and transmits only if its trans-mission does not add noise to the ongoing recep-tion. This scheme as well, avoids the problems dueto asymmetry, since nodes transmit busy tones atthe preset maximum power levels; however, theaforementioned limitations with regards to spatialre-use are also valid for this protocol.

Muqattash and Krunz proposed both dual chan-nel [21] and single channel [22] MAC protocols thatenable power control in ad hoc networks. Theapproaches take into account the potential interfer-ers while administering power control. However, aswith the schemes in [11,20], these protocols requirehigh power transmission of the control packets sothat all potential interferers are reached. Such highpower transmissions might not be possible in themultifarious, inherently power heterogeneous net-works that we consider.

2.2. Link asymmetry due to topology control

Research on topology control in ad hoc networksvia the deployment of multiple transmission powerlevels has particularly focused on (i) minimizingthe overall energy consumption [14,6], (ii) boundingthe node degree [17,15], and (iii) reducing the inter-ference to increase effective network capacity [7].The proposed solutions are based on adjusting thetransmission power to avoid power-intensive com-munications [16,15,32]. Towards this, nodes choosea subset a of their neighbors to directly communi-cate with, and they reduce their transmission power(from the maximum) to the value sufficient to reachthe farthest neighbor in a [16,15,32,6]. Ramanathanproposed a centralized topology control protocol tobound the energy consumption in the network [28];with the proposed approach, nodes reduce theirtransmission powers to the extent possible withoutsacrificing the network connectivity. Similarly, theCOMPOW protocol proposed in [23] requires thenodes to operate at the smallest power level at whichthe network is connected. This power level is com-puted by routing layer agents, and is used by allnodes in the network. Another approach for facili-tating power control is clustering, as proposed in


[13]. With this approach, nodes adjust their powerlevels so as to communicate only with the cluster-head. The above efforts are geared towards topol-ogy control via power adaptations; they do notexamine the effects of the created asymmetry onprotocol performance at various layers.

2.3. Handling link asymmetry at the routing layer

Routing in the presence of unidirectional linkshas been studied in [2,3,19,27,31]. In [19], authorspropose to bypass the unidirectional links and routethe packets via only bi-directional links; this strat-egy may lead to the overloading of certain links,while under-utilizing others. In [3], reverse multi-hop paths are used to proactively discover and useunidirectional links. In [31], the authors extend thewell-known zone routing protocol (ZRP) [10] tosupport unidirectional links. With these routingschemes, at the MAC layer either the traditionalIEEE 802.11 MAC protocol is employed, or anideal MAC protocol that can handle unidirectionallinks is assumed. While the former causes perfor-mance degradation due to unfairness (as mentionedbefore), the latter approach does not accuratelyreflect in the trends in the latter approach is unableto utilize system resources efficiently.

Link layer tunneling approaches to support rout-ing in the presence of unidirectional links have beenexplored in [9,24]. These approaches hide the unidi-rectional nature of a link from higher layer proto-cols so as to facilitate their operations without anymodifications. Tunneling is based on forming areverse multi-hop path for each unidirectional linkusing the information gathered by the routing pro-tocol. The reverse path is also sometimes referredto as the inclusive cycle [3]. A similar idea appearsin [29], wherein a sub-layer beneath the routinglayer is developed. The work also attests to theproblems incurred in routing due to the lack ofproper MAC layer protocols that handle asymme-try. There is also some work on using multi-hopacknowledgements to discover unidirectional links[2,24]. GPS-based approaches for enabling link levelacknowledgements [12] over unidirectional linkshave also been proposed. The above link layerapproaches however ignore the implications of linkasymmetry at the MAC layer. In [19], the impact ofunidirectional links on routing performance is stud-ied and it is once again identified that more efficientMAC protocols are needed to handle unidirectionallinks.

2.4. Study of medium access in power heterogeneous

networks

In [26], Poojary et al. quantify the inefficiencies inthe use of the IEEE 802.11 MAC protocol in a net-work wherein nodes have heterogeneous powercapabilities. A heterogeneous network with twotypes of nodes, operating at power levels of0.56 W and 0.14 W was considered; it was shownthat the low power nodes suffer up to 50% degrada-tion in throughput as compared to the high powernodes, under various conditions of network load.This degradation was a consequence of the factthat the transmissions of low power nodes wereoften interfered with transmissions from highpower nodes that were unable to hear the RTS/CTS exchange between the low power nodes. Theauthors proposed to solve this problem by meansof a CTS propagation technique using a standardflood-type broadcast algorithm. The algorithmrequired that nodes that hear a CTS frame wouldpropagate the frame further (up to a distance deter-mined by the ratio of the range of high powernodes to the range of low power nodes). Thepropagated frames were called the bandwidth reser-vation (BW_RES) frames. The objective of thisbroadcast was to inform the high power nodes inthe neighborhood about the ongoing communica-tion so that they would inhibit their own transmis-sions for the duration specified in the BW_RESframe. It was found that such a flood type broadcastdid not offer performance benefits; in fact, thescheme caused a further degradation in terms ofthroughput since the overhead incurred in propa-gating these BW_RES control frames outweighedthe potential gains in terms of reducing the numberof collisions. This work neglects the presence of‘‘sensing range’’ with the IEEE 802.11 MAC proto-col [11]; furthermore, the effects at higher layershave not been considered.

In [4], Bao et al. propose a MAC protocol for adhoc networks with unidirectional links. The pro-posed approach schedules node transmissions basedon the contention in their one and two-hop neigh-borhoods such that fairness is ensured. The schemehowever, depends on the existence of accurate 2-hopneighborhood information at each node andrequires perfect clock synchronization. Our pro-posed framework is designed to solve the problemsdue to link asymmetry at the MAC layer, withoutthe requirements or constraints that limit the practi-cality of this solution.

1 The problems have been discussed in detail in [26,19,29,3].2 Typically two ranges are defined for the transmissions of a

given node u viz., the transmission range and the interferencerange [1]. Nodes that are within the transmission range of u candecode the frames received from u. Nodes that are within theinterference range but not within the transmission range of u

cannot decode frames from u; however they still interfere with atransmission of u.

3 Lost route update packets can lead to falsified routing tableswhen traditional proactive routing schemes are used. A lowpower node could wrongly assume that it can reach a high powernode and a high power node may not know that it could reach alow power node. We omit the detailed discussion of these effectsin this work since they are discussed in prior papers on routing onunidirectional links.


In [18], Liu et al. propose a three phase approachfor power heterogeneous ad hoc networks. Thethree phases are time division multiplexed. Linkasymmetry is controlled by allowing the heteroge-neous power transmissions only in one of thephases. The scheme imposes that within this phase,at any given time, only a single high power node canbe active. This is because the MAC scheme does notutilize the CSMA/CA capability of the IEEE 802.11MAC; it requires perfect scheduling between highpower transmissions. The paper does not addressthese scheduling issues. This constraint may causean under-utilization of the available bandwidth.In addition, the proposed scheme works efficientlyonly in scenarios with low mobility or long pausedurations.

2.5. Summary

To summarize, the majority of the previousefforts only address the link asymmetry problem atspecific layers. Most of the efforts propose solutionsfor routing via unidirectional links. There are a fewapproaches that study link asymmetry from the per-spective of the MAC layer. However, they are lim-ited in scope. In particular, the distinguishingcontributions of our work in contrast with existingliterature are:

• Most previous approaches that propose solutionsfor routing via unidirectional links ignore MAClayer effects. In particular they either assume anideal MAC scheme or simply use the IEEE802.11 MAC without modifications. Our schemeis the first to consider the interactions betweenthe layers.

• Existing MAC layer solutions for handlinglink asymmetry are based on imposing time-schedules. These schemes require frame/clocksynchronization and in some cases, the use ofhomogeneity within the network. Furthermore,the previous studies do not consider theimpact at higher layers (routing in particular).Our cross-layer approach overcomes theselimitations.

• Similarly, previous power control based mediaaccess schemes typically handle link asymmetryby assuming that all nodes can transmit with afixed maximum transmission power level. Ourapproach is applicable in a truly heterogeneoussetting where the above assumption does nothold.

3. Problem statement

In this section we revisit1 the performance of thedistributed coordination function (DCF) of theIEEE 802.11 MAC protocol in a power heteroge-neous network setting. We discuss its deficienciesand highlight the resulting effects on the higherlayers.

As alluded to earlier, the inefficiency of the IEEE802.11 MAC protocol arises primarily in scenarioswith link asymmetry. Link asymmetry causes lowerpower nodes to be hidden from higher power nodes.This, in turn, increases the number of collisions thatare experienced by low power communications. Thiseffect is depicted in Fig. 1. The RTS/CTS exchangebetween two low power nodes A and B is not over-heard by node H since H is not within the sensing orinterference range2 of these communicating nodes.Thus, it is possible that while the data exchangebetween nodes A and B is in progress, node H couldinitiate another transmission, and cause a collisionat node B.

A second problem that is manifested at the MAClayer occurs when a node fails to identify (and uti-lize) a unidirectional link. This effect is depicted inFig. 2, where H1 can reach L1 but not vice versa.As a result, if L1 responds to any frame (such asan RTS frame) sent by H1, the response never getsto H1. Similarly, any control frame initiation byL1 (e.g. an RTS frame) would never reach H1.Depending on the scenario, these problems couldcause degradations due to wasteful control frametransmissions and back-offs at the MAC layer.The link asymmetry can also degrade the perfor-mance of traditional on-demand routing protocols3

due to the loss of control packets. One such effect isdepicted in Fig. 2 where node H1 attempts to estab-lish a route to H2 through nodes L1 and L2. The

B

C

A

H

RTS

CTS

CTS

BW_RESRange of A

Range of B

Range of H

H2

H1

BW_RES

BW_RES

Fig. 3. CTS propagation scheme.

Range of H2

Range of H1

Range of H

H

H2

H1

L1

L2

Fig. 2. Problem at the MAC/routing layers due to link asym-metry in the power heterogeneous network.

B

C

A

H

RTS

CTS

CTS

Range of A

Range of B

Range of H

Range of C

Fig. 1. Problem at the MAC layer due to link asymmetry in thepower heterogeneous network.


routing control packets from L1 is not received byH1 since it is outside the range of L1. Such effectscould lead to repeated (albeit unsuccessful) routediscovery attempts. We do not discuss these in greatdetail since prior work on unidirectional routingtouch upon such problems [3,19].

4. Our cross-layer framework

In this section, we present our cross-layer frame-work for efficiently handling link asymmetry at theMAC and routing layers. The key idea of ourapproach is the selective forwarding of low powercontrol packets to hidden high power nodes. In thissection, we first introduce the components that aidthe implementation of our cross-layer approach.Next, we study two extensions for a modularMAC layer solution, and discuss their performanceenhancements. Finally, we present our cross-layerapproach.

4.1. Preliminaries

In the following, we first revisit the BandwidthReservation (BW_RES) frame structure introducedin [26]; next, we compute the effective hop count upto which this frame should be forwarded such thatasymmetry is eliminated.

The BW_RES frame: This frame is used to pre-vent high power nodes from initiating transmissionsthat may collide with ongoing lower power commu-nications. The key idea is to have the nodes thathear a low power CTS frame broadcast a copy ofthe frame (the BW_RES frame) to reach nodes thatare more than one low power hop away (Fig. 3).Reception of a BW_RES frame at a high powernode indicates that a lower power communicationis active in its vicinity, and that the node shouldinhibit its transmissions (for the time period speci-fied in the BW_RES frame).

We modify the BW_RES and CTS frame struc-tures to aid the realization of the proposed perfor-mance enhancements; the new frame structuresand a timing diagram of the modified reservationscheme depicting the BW_RES transmissions areshown in Fig. 4a and b. The BW_RES frame formatis similar to the RTS frame format except thatthe frame control field has a few additional attri-butes: (i) a sequence number, and (ii) the originatoraddress field that contains the MAC address of theRTS sender. Our framework uses these fields fordetecting duplicate BW_RES frames that may bereceived by third party nodes. A sequence numberfield is similarly added to the CTS frame structure.In the ‘‘frame control field’’ of CTS/BW_RESframes, we use the To DS, From DS and MORE bits

2bytes 2bytes 6bytes 2bytes4bytes6bytes

Seq NoOriginatorAddress

DestAddress

DurationFrameControl

FCS

6bytes

BW_RES

2bytes 2bytes 2bytes4bytes

Seq No Dest AddressDurationFrameControl

FCS

CTS

RTS

BW_RES

DATA

CTS

BW_RES

ACK

DATA

ACK

Multiple Reservation Scheme

SIFS SIFS6*SIFS

6*SIFS

SIFS

Neighbor

Neighbor

Receiver

Sender

SIFS SIFS

Fig. 4. Modifications on the IEEE 802.11 MAC protocol and on the scheme proposed in [26] to offer performance enhancements.(a) Modified frame structures. (b) Modified reservation scheme.

4 This approach is similar to the IEEE 802.11 MAC protocol,where, upon the collision of RTS and/or CTS messages, nodessimply use physical carrier sensing to set their NAVs for anextended inter-frame space period.


(used for signaling in infrastructured wireless net-works in IEEE 802.11 [1]) to indicate a TTL valuefor the BW_RES frames. In our experiments weuse TTL = 2; the choice for this value is justifiedin the following. With our cross-layer framework,the initiator of the CTS frame chooses the set ofnodes (say L in number; L is typically small) thatare to forward the BW_RES frame; the list of theseL nodes is appended into the CTS and BW_RESframes. In our simulations we account for the over-head incurred due to this additional information inthese frames.

Nodes that overhear the BW_RES frames settheir network allocation vectors (NAVs) appropri-ately (the duration of the communication is indi-cated). If a node simply senses energy due to oneor multiple BW_RES frames and is unable todecode a BW_RES frame, it simply sets its NAVto indicate that the medium is busy for an extendedinter-frame space (EIFS) interval as with traditionalIEEE 802.11.

We remark that the BW_RES frame is as small asthe CTS control frame, except that it carries the IDsof the nodes that are to forward this frame. Hence,the collision of these frames has a small likeli-hood. Furthermore, nodes perform carrier sensingbefore transmitting BW_RESs; this further reducesthe probability of a collision. However, if theBW_RESs collide, a subset of the high power nodesmay not hear them; they will however sense thechannel to be busy and therefore set their NAVsfor an EIFS period.4 Still, these nodes may initiatetransmissions that collide with low power communi-cations; in our simulations we find that these effectsare not pronounced.

In the following, we determine the appropriatedistance (in terms of hop count) up to which the


BW_RES frames should be propagated. Specifi-cally, we justify the selection of value 2 for theTTL of BW_RES frames.

At the physical layer, we adopt the path losschannel model from literature. With this model, ifPtx is the transmission power of a node u, thenthe received power at a distance d from u is givenby Prx = Ptx Æ d�a, where a is the path loss exponentdepending on the wireless transmission environ-ment. We denote the power levels of high powerand low power nodes as Pmax and Pmin, and the cor-responding transmission ranges as dmax and dmin,respectively.

The ad hoc nodes define a unit disk graph(UDG), wherein a link exists between two nodes ifand only if they are within a unit distance from eachother, where the maximum transmission range ismapped to unit distance. While it is known thatthe transmission range of a node can be affectedby wireless channel impairments such as multi-pathfading and shadowing, we consider only the pathloss to simplify our analysis. According to thismodel, the signal-to-noise ratio (SNR) of a framereceived at the periphery of a unit disk is the thresh-

old value that is necessary for the correct decodingof a received frame. This SNR threshold is the sameat both high and low power nodes. Our objective isto ensure that the CTS frame of a low power node ureaches all high power nodes that have u withintheir transmission range. Given the UDG model,we note that:

P max

P min

¼ dkmax

dkmin

and hence;dmax

dmin

¼ P max

P min

� �1=k

: ð1Þ

For the chosen5 values of Pmax, Pmin and k weobtain dmax = 2 Æ dmin. Thus, if the BW_RES frameof a low power node u is propagated through twolow power hops, the frame can with high probabil-

ity reach the high power nodes whose transmissionis received at u. We may refine our model to dis-tinguish between the sensing and the correct recep-tion of a packet (in the former the packet isreceived but cannot be correctly decoded due to

5 In our simulations we chose Pmax = 0.56 W, Pmin = 0.14 W,and k = 2, representing propagation in free space. These valuesare parameters input to the simulations and can be altered fordifferent scenarios. The behavioral results that we observed (viasample runs) with other scenarios were similar to the onesreported here.

its low SNR value). In literature, the sensing rangeis typically modeled to be twice the transmissionrange; the frame must be propagated through fourlow power hops (or respectively two high powerhops, from (1)). Choosing a TTL value of 2 is seen(in our simulations) to provide the requisite bene-fits; however, we remark that it does not guarantee

that a BW_RES frame will reach all high powernodes that may potentially initiate colliding trans-missions (high power nodes may be more than 2hops away from a low power node). Yet, wesupport with our simulation results in varioussce- narios that it is more effective to restrain thepropagation to a localized vicinity of the commu-nication as opposed to invoking a broadcast witha wider scope.

4.2. Enhancements at the MAC layer

Before designing our cross-layer framework, weconsider simple modifications to the IEEE 802.11MAC protocol to enhance the MAC layer per-formance in power heterogeneous networks. Inthis subsection, we describe these modifications(namely the multi-reservation and the intelligentbroadcasting techniques) and explain our intu-itions for making the reported design decisions.The proposed approaches offer enhancements atthe MAC layer, but introduce an overhead tothe ad hoc network. Furthermore, these ‘‘MAC

layer enhancements’’ do not provide support forrouting over unidirectional links. We address theseproblems in the next subsection, where we proposeour ‘‘cross-layer framework’’, which provides meth-ods to handle asymmetry at the routing layer aswell.

4.2.1. Reserving bandwidth for multiple sequential

transmissions

As discussed earlier, [26] considered the scopedflooding of the BW_RES frame to preclude highpower transmissions in the vicinity of ongoinglow power transmissions; however, the overheadincurred was prohibitive. To limit the overheaddue to broadcasts of BW_RES frames, we attemptto reduce their frequency. One way of achieving thisis using a single RTS/CTS/BW_RES initiation formultiple sequential DATA/ACK exchanges. Themultiple DATA frames can be independent, i.e. theymay have separate fields, including the checksum.Before sending an RTS, a node checks its interfacequeue (between the network and the MAC layers)


for other DATA frames with the same destinationaddress as that of the RTS. If such frames arefound, they are moved to the MAC layer and buf-fered along with the original (RTS) frame to be sent.The RTS frame attempts to reserve the channel fora preset number (say N) of data frames at a time.When a node does not have additional frames forthe destination (i.e., there is only one frame thatcan be transported at the given time), it reservesthe channel for the single data frame.

Clearly, allowing a node to reserve bandwidth fora large number of sequential transmissions couldlead to unfairness in the network. On the otherhand, if only a small number of transmissions arereserved with a single control frame, a significantreduction in the number of generated BW_RESframes may not be feasible. We find by simulationsthat reserving the channel for 2 and 3 sequentialtransmissions provides a significant reduction inthe volume of broadcast BW_RES frames. In ourfuture discussions we refer to this technique as themulti-reservation technique. The sequence of frametransmissions due to the multi-reservation techniqueis depicted in Fig. 4b.

4.2.2. Intelligent broadcasting

To further reduce the volume of broadcastBW_RES frames we examine the benefits of exclud-ing unnecessary re-broadcasts. Our approach isderived from a broadcast scheme proposed in [25],and it executes as follows. Each node u, upon receiv-ing a BW_RES frame, sets a randomly chosen time-out in the future. It then records (in a counter) thenumber of subsequent BW_RES broadcasts (all ofwhich correspond to the same CTS frame) that itoverhears via broadcasts from its neighbors priorto the time-out. If the value indicated by the counterexceeds a preset fixed threshold T (a tunable systemparameter), u revokes its own BW_RES broadcast.The idea behind this scheme is that overhearingmultiple copies of the same broadcast frame sug-gests a node that it is located in a highly denseneighborhood, which implies that the additionalcoverage it would achieve by performing its ownbroadcast is likely to be insignificant. We find bysimulations that a choice of T = 3 preserves cover-age and significantly reduces the number of unnec-essary broadcasts. We observe that at lower valuesof T there is a reduction in coverage (nodes quelltheir transmissions to a large extent), and for highervalues there are broadcasts that do not offer anenhancement in coverage.

4.2.3. Failures and retransmissions

The actions taken upon transmission failure arethe same as performed in the IEEE 802.11 standard;if the RTS sender does not receive a CTS response,it backs off and attempts to retransmit the packetafter a back-off period. As with the IEEE 802.11MAC specifications, we impose an upper boundon the number of retransmission attempts for RTSand DATA packets (7 and 4, respectively). Succes-sive packets after these pre-specified number ofattempts are dropped. If a sender does not receivean ACK for a specific DATA packet among themultiple sequentially attempted transmissions (withour multiple reservation scheme), it attempts toretransmit only the particular failed DATA packetafter backing-off. During the retransmission of thispacket, the sender will again attempt to reservebandwidth for additional packets destined for thesame node.

4.3. Integrated MAC/routing framework

In the following, we describe our cross-layer

framework for supporting medium access controland routing in power heterogeneous ad hoc net-works. Specifically, our cross-layer frameworkimproves the enhancement achieved at the MAClayer by the mechanisms introduced in the previoussubsection; in addition, it introduces methods tohandle the asymmetry at the routing layer. In ourframework, the MAC layer solicits assistance fromthe routing layer in determining a small subset ofnodes that will perform the BW_RES re-broadcasts.In return, the routing layer depends on the MAClayer for the discovery and use of unidirectionallinks. Thus, the key functionalities of our frame-work are the intelligent propagation of BW_RESmessages and the construction of reverse routesfor bridging unidirectional links.

4.3.1. Topology-aware CTS propagation (TACP)

We propose a routing-assisted approach forreducing the number of propagated BW_RESframes. With this approach, nodes multicast theBW_RES frame to the high power nodes in theirvicinity, as opposed to broadcasting this frame.

To facilitate this, we require that each nodemaintains link-state information with regards toits two-hop neighborhood [8]. This information iscollected via Hello messages as follows. At network


instantiation, every node broadcasts6 a list of itsone-hop neighbors (we call these one-hop neighborsof a node u, as the in-bound neighbors of u) that it iscurrently receiving Hello messages from. TheseHello messages also contain the corresponding max-imum transmission power (in watts) for each in-bound neighbor included. After the initial phaseeach node constructs an inbound tree with the Hello

messages it has received. The inbound tree of a nodeu includes all neighbors that can reach node u. Asthe network reaches a steady state, nodes begintransmitting Update Hello messages that are nowmodified to contain their inbound trees. These mes-sages are broadcast every ‘‘Hello interval’’ millisec-onds. Each node then combines the in-bound trees

reported by its neighbors with that of its own, andforms a localized graph that depicts its local neigh-borhood. The Update Hello messages are furthermodified to include this localized neighborhood.Periodic transmissions of the Update Hello messageshelp refine this localized graph.7 As informationpropagates, the localized graphs become moreextensive. This allows nodes to gather additionalinformation (beyond their two hop neighborhoods);however, the sizes of Update Hello messages growconsiderably. There is a trade-off between theamount of information propagated and the extentof knowledge that is possessed by a node withregards to its vicinity. In our studies, nodes simplyprune nodes that are beyond a certain number (n)of hops from their localized graphs; and this prunedgraph is included in the Update Hello messages. Ineach update, by transmitting only the changes tothe localized graph as compared with the previousupdate, one may significantly constrain the size ofthe Update Hello messages. As mentioned earlier,a choice of n between 2 and 4 ensures to a greatextent that most of the high power nodes that affecta given low power communication are informed bymeans of BW_RES frames. Our simulations suggestthat setting n = 3 offers the best benefits8; our sam-ple studies suggest that with n = 2, a significant frac-tion of the high power nodes are missed by the

6 Hello messages have a TTL (time to live) value of 1, i.e. theyare only exchanged between one-hop neighbors.

7 The periodicity of the Update Hello messages would dependon the mobility in a given scenario. If nodes are highly mobile,the Update Hello messages must be transmitted with higherfrequencies.

8 This implies that all high power nodes that can be reached viathree low power hops from a low power communication areinformed of the impending communication.

BW_RES broadcasts due to collision effects whilen = 4 does not offer extended coverage benefits.We justify the chosen value for this parameter,based on our simulation results in Section 5.

Using the localized graph, our objective is then tohave each node construct an n-hop outbound Steinertree on which the BW_RES frames will be multicast

to the high power nodes. Note that a low powernode u does not initiate the propagation of theBW_RES frame if there are no high power nodesin its n-hop neighborhood. If high power nodes existwithin the n-hop neighborhood of u, u identifies theminimum set of nodes in its one hop neighborhoodthat can reach all other high power nodes in itstwo-hop neighborhood. We refer to these nodes as‘‘Candidate Nodes’’ for relaying the BW_RESframe. Node u includes the IDs of these nodes inits CTS frame. As our goal is to minimize the num-ber of BW_RES re-broadcasts and to reduce (to theextent possible) the latency incurred during a MAClayer exchange, the candidate nodes are typicallychosen to be high power nodes (if such nodes areavailable). The one-hop relays then perform a simi-lar computation to identify the next set of relays(if needed). The IDs of this next set of candidaterelays are included in the BW_RES frame. If a node,upon receipt of either a CTS or a BW_RES framedoes not find its ID in the frame, it simply updatesits NAV (network allocation vector) in accordancewith the IEEE 802.11 MAC protocol and discardsthe frame. The multi-reservation scheme is incorpo-rated as well; if possible, a node reserves the channelfor N data frames destined for the same neighbor bymeans of a single RTS/CTS/BW_RES initiation.Since the BW_RES frame is multicast to the opti-mum set of one-hop neighbors along the node’sSteiner tree (as opposed to simple broadcasting),the number of propagated BW_RES frames doesnot increase with increasing local density. In fact,a better set of forwarding nodes could be viable ata higher density and this would further decreasethe BW_RES overhead. We provide an algorithmicrepresentation of the topology-aware propagationscheme in Fig. 5.

4.3.2. Construction of reverse routes for bridging

unidirectional links

Our framework also proposes a cross-layerapproach, to assist routing in the existence of linkasymmetry. The key idea is to construct reverseroutes that span the unidirectional links. In the

In order to determine the candidate nodes we assume that the following sets have alreadybeen computed by a node using periodic hello messages. These sets are inputs to theprocedure Compute_Potential_Candidate which return a List F consisting of candidatenodes.

Hx1 = Set of all high power nodes in one hop neighborhood of x.Hx2 = Set of all the high power nodes that are 2 hop neighbors of x and not in Hx1.Lx1 = Set of all low power nodes in the one hop neighborhood of x.LHx2 = Set of low power nodes in Lx1 that have high power nodes in their neighborhood.

A node executing the procedure Compute_Potential_Candidates should satisfy followingconditions

1. It should have its identifier in the received CTS or the BW_RES message.2. The TTL field in the received message should be greater than zero.

If above conditions are not met then the node just updates its NAV as specified by thestandard legacy IEEE 802.11 MAC and does not take any further action. Let S be thenodes that hear the CTS or BW_RES message and let the indicated TTL be higher thanzero.

Procedure Compute_Potential_Candidates (Hx1 , Hx2 , Lx1 , LHx2){

For each node x in S Do

Add Hx1 to FIf

Hx1 = nil and LHx2! =nilThen add LHx2 to F Else

I f ( LHx2! = nil )Then add to F those nodes of LHx2 that can reach high power nodes not inHx1 and not reachable by nodes in Hx1.Else

if (LHx2 = nil and Hx1 = nil)F = nil (if node is replying to a RTS then just broadcast CTSmessage as in 802.11 MAC signaling else do not propagate CTSor BW_RES further)update NAV

return F }

Fig. 5. Algorithm to select potential candidates.

L1

H1

Range of H2

Range of H1

H2L2

Range of L1

Range of L2

Fig. 6. Reverse route construction to route MAC/routing controlpackets.


following, we describe how these reverse routes areformed.

The exchange of the previously discussed Hello

messages help in detecting the unidirectional linksin the network. A node detects that it is at the tail

of a unidirectional link, when it receives a Hello

message from a neighbor and finds that it isexcluded from this neighbor’s ‘‘neighborhood list’’.Unidirectional links are also depicted in nodes’pruned localized graphs. Using this graph, a nodethat is at the tail of a unidirectional link can com-pute a reverse path to the node at the head of thisunidirectional link. We illustrate this process inFig. 6. In this figure, the link between nodes L1

and H1 is unidirectional (H1 can reach L1 but L1

cannot reach H1). To utilize this link, L1 constructs


a reverse route to H1. L1 learns from L2’s inboundtree (using the method explained above) that L2

can reach H2 which can in turn reach H1. Thusthe reverse path from L1 to the high power nodeH1 is via nodes L2 and H2. If a reverse path fromL1 to H1 of less than n hops does not exist, theunidirectional link (H1,L1) will not be utilized. Itmay be possible to find longer reverse paths;however, discovering such paths would entail signif-icant additional overhead, and thus could actuallyoutweigh the gains incurred in utilizing the link.Once the reverse path is found, L1 encapsulatesthe control frames from the MAC and routinglayers into an IP packet and routes (or tunnels) thispacket using the constructed reverse path. Thisproactive form of routing is only used within thenode’s n-hop neighborhood and any traditionalon-demand routing protocol can then be deployedfor network-wide routing. If either the reverse pathor the unidirectional link were to fail, the tunneledbi-directional link would break. The network-wideon-demand routing protocol would then instigatea route error packet.

With our scheme, Hello messages are tunneled tothe node at the head of the unidirectional link, inorder to inform this node of the existence of thelink. We also tunnel the MAC layer control frames(namely the CTS and ACK frames) and the routinglayer control packets (namely the RREP packet, asproposed in [30]). The motivation for tunnelingthese frames/packets is, that significant benefits areachieved when the unidirectional link is utilized(thereby potentially avoiding a longer alternatepath).

In order to distinguish between tunneled MAClayer frames and network layer packets, the packetheader at the network layer is modified to supporta flag indicating whether a packet is of the encapsu-lated type. At the network layer, this flag can beadded as an option to the IP header, beyond the20 byte standard header. The value of the flag wouldfurther indicate whether the encapsulated packetcontains a MAC layer frame or a routing layerpacket. Upon stripping the outer header, based onthis value, the network layer either forwardsthe packet to the routing module or to the MAClayer.

Finally, we note that our tunneling scheme issimilar in spirit to those proposed in [9,12]. Thekey difference is, that our integrated framework(TACP) for handling MAC layer asymmetryprovides us with a simple and seamless way of

discovering reverse routes to bridge unidirectionallinks.

5. Performance evaluation

We have performed simulations using the eventdriven network simulator ns2 (v.26). We divide theevaluation of our proposed enhancements into twoparts. First, we exclusively measure the enhance-ments at the MAC layer (as proposed in Section4.3) and quantify the benefits as compared to theperformance of the IEEE 802.11 MAC protocoland also the performance reported in [26]. Second,we evaluate the performance enhancements offeredby the implementation of our framework at thehigher layers. We consider various scenarios anddiscuss the observed results.

5.1. Performance evaluation at the MAC layer

5.1.1. Simulation setup

Towards our first goal, to decouple the effects ofrouting and transport layer artifacts from theMAC-layer throughput in this study, we introducea Poisson traffic generation agent above the MAClayer. The traffic generation rate is 1000 packets/s(unless stated otherwise), each data packet being1000 bytes in size. We also vary the average packetgeneration rate in order to observe the effects withdifferent system loads. When a data packet is gener-ated at a node, it is randomly destined for one of thenode’s neighbors.

The network consists of 40 nodes deployed in asquare region. We vary the area of this region inorder to consider different network densities. Theratio of high power to low power nodes is 1, i.e.50% of the nodes are high power nodes and 50%are low power nodes. This choice is based on theresults in [19], which suggest that in the network amaximal number of unidirectional links exist forthis value. We also studied various other parametersettings; however the observed results and the inter-pretations thereof are similar to those reported andthus, are not included.

The physical layer is based on the IEEE 802.11specifications. Nodes move in accordance to a mod-ified version of the random waypoint mobility modelwith a constant speed of 6 mps, and pause for 0.1 sat each destination point on the plane. This constantspeed is chosen in light of the recent results suggest-ing that a choice of random speeds is not appro-priate for a realistic modeling of mobility [34].

Table 1Simulation setup for evaluation at the MAC layer

Simulator ns2 (version 2.26)Number of nodes

(high power/low power)40 (20/20)

Length of square grid Varied from 300to 2000 m

Power levels 0.56 W and 0.14 WTraffic model Poisson, 1000 packets/sMobility model Modified random waypointNode speed, Pause time 6 m/s, 0.1 s


Depending on the speed simulated, nodes choose afrequency between 0.3 and 0.5 s, and generate Hello

messages with this period. The parameters9 used forthis first set of simulations are tabulated inTable 1.

We assume that the communication channel isobstacle-free and that signal degradation occursonly due to path loss (as in other previous work[26,11,20]). We also assume that the channel is sym-metric, and asymmetry occurs due to differences intransmission power levels. The channel bandwidthis set to 2 Mbps. All MAC control frames are trans-mitted at 1 Mbps and data at 2 Mbps, so as to con-form to the IEEE 802.11 standards [1]. We find thatreserving the bandwidth for a maximum of N = 2(for high power nodes) and N = 3 (for low powernodes) sequential transmissions provide significantbenefits.

5.1.2. Parameters

We vary the node density and the traffic load, andwe observe the effects on the performance, in termsof our metrics that we define below.

5.1.3. Metrics

The primary metrics of our interest in this firstpart of simulations are:

• Data success rate. The percentage of successfuldata frame transmissions given the RTS/CTSexchange between the pair of communicatingnodes is successful.

• Throughput efficiency per node. We define thismetric to quantify channel usage; it is the ratio

9 To draw a fair comparison we try to be consistent with thescenario in [26]; this dictates our choice of number of nodes andtransmission power levels, as shown in Table 1.

of the time spent by a node in successfully trans-mitting data, to the total simulation time.

Using these metrics, we measure and compare theperformance of the following:

1. Intelligent broadcasting (IB) and multi-reserva-tions (MR)• Case (a). The legacy IEEE 802.11 MAC

protocol.,• Case (b). 802.11 MAC protocol with

BW_RES propagation using IB,• Case (c). 802.11 MAC protocol using MR,• Case (d). 802.11 MAC protocol with

BW_RES propagation using IB and MR.To observe the performance of the above fourcases on the low power and high powernodes, we define Variant I and Variant II,which stand for (i) only the low power nodesperform the intelligent BW_RES propagationand multi-reservations when initiating a com-munication, and (ii) all nodes, regardless oftheir power level, use these enhancements,respectively.

2. Topology-aware CTS propagation scheme(TACP)• Case (e). The IEEE 802.11 MAC protocol

with BW_RES propagation via TACP,• Case (f). The IEEE 802.11 MAC protocol

with BW_RES propagation using TACP and

MR.

5.1.4. Simulation results

First, we examine the performance of the simpleMAC layer enhancements (Case (d)) with bothVariant I and Variant II, and compare their per-formances with that of the legacy IEEE 802.11MAC protocol (Case (a)). The data success rateof low power nodes with both variants is depictedin Fig. 7a. As compared to scenarios with thelegacy IEEE 802.11 MAC protocol, the low powernodes see an overall improvement of up to 14%with Variant I (and up to 12% with Variant II)with our schemes. Clearly, low power nodes benefitmore when only these nodes use the proposed tech-niques in the network (Variant I); the data successrate for the low power nodes is better by about 2%with Variant I than with Variant II. However, notefrom Fig. 7b that the overall data success rate isbetter with Variant II than with Variant I. Thisis because with Variant II, high power nodes canalso benefit from our schemes; specifically, the

Fig. 8. Link failures in the heterogeneous network.

70

75

80

85

90

95

100

0 500 1000 1500 2000 2500

Length of Square Grid

Dat

a Su

cces

s R

atio

(%)

IEEE 802.11 Variant II Variant I

84

86

88

90

92

94

96

98

100

0 500 1000 1500 2000 2500

Length of Square Grid

Dat

a Su

cces

s R

ate(

%)

IEEE 802.11 Variant II Variant I

Fig. 7. Data success rate in the heterogeneous network with MAC layer enhancements. (a) Data success rate of low power nodes.(b) Overall data success rate.


effects of false link failures10 are alleviated. Asshown in Fig. 8, our proposed framework is ableto reduce the number of false link failures by upto 20%.

We observe in Fig. 7a and b that the data successrates with the raw IEEE 802.11 MAC protocol andwith the MAC layer enhancements are high, both atlow and at high densities; however, there is a degra-dataion observed at moderate node densities. Atlow densities, the probability of a low powercommunication being interfered by a high powercommunication is small (the probability of havinga high power node in its vicinity is small). At high

10 False link failures occur if a node deems that a link has failedas a consequence of successive failed RTS transmission attempts.This happens if the receiver is being interfered with some othertransmission.

densities, the nodes are close to each other andhence the effects of link asymmetry are not severe.Therefore, the performance in these two extreme


cases is fairly good. At moderate densities, the linkasymmetry effects are more pronounced, as theprobability of a low power communication beingin the vicinity of a high power transmitter is notinsignificantly low. Hence the data success rate islow in this regime. The MAC layer enhancementsdo alleviate the degradation but cannot completelyeliminate the effect.

In the same setting we also deployed a homoge-neous network where all nodes have the same trans-mission power of 0.28 W; we observed an increaseof up to 10% in network throughput, and a reduc-tion in the number of false link failures by about15% as compared with the IEEE 802.11 MAC pro-tocol. In sum, the modifications that we consider areshown to provide benefits at the MAC layer in bothpower heterogeneous and power homogeneous net-works. Depending on the percentage of low powernodes in the network one might prefer to use Vari-ant I or Variant II. The former would provide betterperformance if the fraction of low power nodes inthe network is large. Since the number of low powernodes may typically vary in the scenarios consid-ered, Variant II would be the preferred designspecification.

The throughput efficiency offered by Case (d) –Variant I is shown in Fig. 9 for different node den-sities. We observe that the throughput of low powernodes improves by as much as 14% as compared tothe IEEE 802.11 protocol. High power nodes alsobenefit with our enhancements, since reducing thenumber of retransmissions of low power nodesdecreases the overall contention for the wirelessmedium. Furthermore, the aforementioned effectsof false link failures are alleviated. In sum, weobserve an overall improvement in the networkthroughput of up to 12% as compared to the IEEE

0

2

4

6

8

10

12

14

16

200 400 600 800 1000 1200 1400 1600 1800 2000Length of Square Grid

% Im

prov

emen

t ove

r IEE

E 80

2.11

Low Power (0.14W) node High Power (0.56W) node

Fig. 9. Increase in average throughput with the MAC layerenhancements.

802.11 MAC protocol. Fig. 9 indicates that at highnode densities (when the area of deployment issmall) the performance improvement with ourscheme is not significant. As alluded to earlier, thisis because, even with a low power (of 0.14 W) thetransmission range is about 205 m, which impliesthat all nodes in the deployment area are withinthe sensing range of each other, i.e., no asymmetryexists. At lower densities, however, link asymmetryin the network increases and it is when the benefitsof our schemes become significant. The networkbecomes highly sparse when the size of the deploy-ment area is further increased. This reduces the pos-sibility of collisions; this case does not pose asignificant challenge on the MAC throughput andtherefore the gains of our scheme are again notvisible.

We observe that the frequency of multi-reserva-tions increases with load (Fig. 10), since at higherloads when a node wishes to initiate a transmission,it is more likely that the node finds multiple packetsdestined for the same neighbor. Thus the efficiencyof the scheme improves with load: the improvementin throughput is about 5% at a load of 500 packets/s,whereas with a load of 1000 packets/s, it is about12%.

In the following experiments, we evaluate theperformance of TACP exclusively (Case (e)) andin conjunction with the MR scheme (Case (f)),and compare this with the performance with Case(b) and Case (d).

We first compare the data success rate with Case(f), with that of the legacy IEEE 802.11 MACprotocol (Case (a)) for low power and high powernodes at different node densities. As shown inFig. 11, in the power heterogeneous scenario thelow power nodes see a significant improvement (of

0

10

20

30

40

50

60

70

0 200 400 600 800 1000 1200Load (Packets/ Second)

% o

f Tim

es M

uli-R

eser

vatio

n O

ccur

s

Fig. 10. Frequency of multi-reservations at various traffic load.

74

79

84

89

94

99

0 500 1000 1500 2000 2500Length of Square Grid (meters)

Dat

a Su

cces

s R

ate

(%)

IEEE 802.11 MAC Homogeneous Network- High Power NodeIEEE 802.11 Homogeneous Network- Low Power NodeIEEE 802.11 MAC Heterogeneous -HighPowerNodeTACP Heterogeneous Network -Low Power NodeHeterogeneous Network IEEE 802.11 MAC- Low Power Node

Fig. 11. Data success rate in the heterogeneous network withTACP scheme.

02468

1012141618

0 500 1000 1500 2000 2500Length of Square Grid (meters)

No

of n

odes

Counter Based Broadcasting Standard Flooding Algorithm TACP

trans

mitt

ing

BW_R

ES

Fig. 12. BW_RES overhead with flooding, intelligent broadcast-ing and TACP schemes.

Fig. 13. Increase in average throughput with TACP in powerheterogeneous settings. (a) Throughput of nodes in a heteroge-neous network. (b) Throughput of nodes in a network with threepower levels.


up to 20%) with Case (f) as compared to the sce-nario with Case (a); their performance is almost asgood as that of the high power nodes. In addition,the number of false link failures decreases withTACP by about 28% compared to the IEEE802.11, and by about 8% compared to when Intelli-gent Broadcasting is used (Fig. 7). The latterenhancement is due to the additional intelligenceat the MAC layer, which significantly reduces theoverhead traffic in the network and the contentionfor channel access. To elucidate this further, wecompare the number of BW_RES frames generatedper CTS instantiation with the standard floodingscheme, Intelligent Broadcasting and TACP. Thenumber of BW_RES frames broadcast per CTSframe reduces by about 50% with TACP even ascompared to that with IB (Fig. 12). In addition,the overall improvement (relative to the IEEE802.11 MAC protocol) in terms of data success rateand throughput are higher with TACP than with IBdespite the overhead incurred due to the Hello mes-sages.11 We observe that with this improvement indata success rate (Fig. 11) and the reduction ininterference from high power nodes, the throughputefficiency of the low power nodes increases by up to24% (Fig. 13a). TACP also alleviates false link fail-ures at high power nodes, this is reflected by theimprovement of up to 12% in their throughputefficiency.

Up to this point, we assumed two power levels –high and low – in our simulations. Next, we intro-

11 TACP does require the transmission of Hello messages unlikethe Intelligent Broadcasting scheme.

duce additional heterogeneity by incorporating mul-tiple possible power levels. The fraction of nodesbelonging to each power level is (almost with negli-gible variation) equal. In Fig. 13b, we depict thethroughput efficiency improvement for the nodes


with the three power levels in the network. Weobserve that our scheme enables the lower powernodes to have a fair share of channel bandwidth.Overall improvement in network throughput ascompared to the IEEE 802.11 MAC protocol is upto 18%. Again, the improvements are more visibleas we increase the size of the deployment area, sincethe asymmetry is emphasized at moderate densitiesas opposed to extremely high densities as describedearlier.

We also quantify the impact on overall MACthroughput, of TACP and IB techniques in isolationfrom the MR scheme. Specifically, we simulate Case(b), Case (c), Case (e) and Case (f); Fig. 14a com-pares Case (b) and Case (c), and Fig. 14b comparescases (e) and (f). While the mere deployment of IBdoes not result in drastic improvements in the over-all network throughput, MR scheme by itself offersan improvement of up to 9%. However, merelydeploying TACP improves performance as much

Fig. 14. Throughput improvement using IB and TACP alongwith multi-reservation scheme, Case (b), Case (c), Case (e) andCase (f). (a) Improvements with the intelligent broadcasting. (b)Improvements with the TACP.

as 16% and MR provides an additional improve-ment of up to 8%. In sum, TACP and IB functionindependent from MR and can be used in conjunc-tion to offer supplementary benefits.

5.2. Performance evaluation at the higher layers

In this second part of our simulations, our objec-tive is to incorporate the routing and transportlayers atop our MAC layer, and quantify the perfor-mance at higher layers.

We now consider UDP traffic with constant bitrate (CBR), and evaluate the performance of twoof the most commonly used on-demand routingprotocols AODV and DSR. We assume that eachnode is represented by its IPv4 address (32 bits),and that the power level of a node is representedby a single bit to indicate whether the relevant nodeis a high power node or a low power node.12 Thesimulation parameters used in these experimentsare listed in Table 2.

We are interested in quantifying the performancein terms of the following metrics:

• Packet delivery ratio (PDR): Ratio of the num-ber of higher layer packets delivered, to the num-ber of such packets generated.

• Average end-to-end delay: Mean end to end delayexperienced by the packets.

• Route search attempts: Number of initiated routediscovery attempts.

• Route search failures: Number of times that asource node fails to find a path to its desti-nation.13

Fig. 15 depicts the performance improvement atthe routing layer by using our cross-layer frame-work at different levels of heterogeneity, as com-pared to using the 802.11 MAC protocol. Thelevel of heterogeneity refers to the percentage oflow power nodes in the network.14 We first study

12 If multiple power levels are to be used, the number of bits thatspecify the power level need to be increased in a logarithmicproportion.13 A route discovery may fail due to two reasons: (i) the network

is disconnected and the source and destination belong to differentnetwork partitions; (ii) there are unidirectional links on the pathfrom the source to the destination. While in the first case none ofthe schemes can compute a route, in the second case ourframework can discover routes while the traditional methods fail.14 This value is maximized when the fraction of low power nodes

is almost equal to that of high power nodes in the network [19].

Table 2Simulation setup for evaluation at the higher layers

Simulator ns2 (version 2.26)Channel bit rate 2 MbpsRadio model Lucent WaveLANNumber of nodes 80Heterogeneity 10–50% low power nodesPower levels 0.56 W and 0.14 WTraffic Model CBR, 5–50 packets/s, 512 bytes/packetNumber of source

nodesVaried between 10 and 15

MAC protocols IEEE 802.11 MAC DCFIEEE 802.11 w/MAC enhancementsIEEE 802.11 w/cross-layer framework

Routing protocols AODV, DSR, AODV/DSR w/cross-layerframework

05

1015202530

10 20 30 40 50Level of Heterogenity (%)

% Im

prov

emen

t in

PDR

MAC Enhancements(AODV)Cross-Layer Framework (AODV)MAC Enhancements(DSR)Cross-Layer Framework (DSR)

05

101520

25303540

0 10 20 30 40 50Levelof Heterogeneity (%)

% D

ecre

men

tin n

umbe

r of r

oute

disc

over

ies

AODV DSR

0

5

10

15

20

25

30

35

1 2 3 4 5 6Level of Heterogeneity (%)

% D

ecre

men

t in

rout

e se

rach

fa

ilure

s

DSR AODV

Fig. 15. Performance improvements at the routing layer with ourcross-layer framework. (a) Increase in packet delivery ratio(PDR) of nodes in the heterogeneous network. (b) Percentagedecrement in route discovery attempts. (c) Percentage decrementin route search failures.


the performance in terms of the first metric, PDR.Fig. 15a depicts that our MAC layer enhancementsregister PDR improvements of up to 12% over thetraditional IEEE 802.11 MAC protocol. The PDRfurther increases with our cross-layer frameworkwith both AODV and DSR, due to the ability toidentify and utilize unidirectional links. Theimprovement is higher with AODV than withDSR at higher levels of heterogenity. This is becauseAODV, with its traditional settings, does not sup-port asymmetric links. DSR, on the other hand,allows a destination (upon the receipt of a routerequest) to invoke its own route discovery to dis-cover the source in the presence of unidirectionallinks. Therefore, AODV has more to gain whendeployed over our framework. We remark thatDSR benefits from our framework as well, sinceits reverse route discovery floods are no longerneeded. These improvements are owing to the over-all reduction in network contention with decreasingMAC and routing layer control packet overhead.This is borne out by Fig. 15b, which depicts thatthe number of route discovery attempts are signifi-cantly reduced (by about 35% for AODV and byabout 25% for DSR) with our framework. Thisreduction is a consequence of our frameworkenabling nodes to easily discover and use unidirec-tional links. Without our framework these linkswere either rendered useless (in the case of AODV)or were discovered with a high overhead (withDSR). Our results also show that the percentageof route search failures are reduced by up to 25%;Fig. 15c depicts the performance for both routingprotocols under consideration.

Finally, we study the mean end-to-end delay atthe routing layer, experienced by packets under sce-narios with different levels of heterogeneity. Fig. 16a

shows that the mean delay experienced is onlymarginally increased by our cross-layer framework.Conflicting factors contribute towards these results.On the one hand, our framework requires the trans-mission of additional BW_RES frames at the MAClayer and thus, a MAC layer transmission takes alonger time than with the traditional schemes. Onthe other hand, this effect is somewhat offset viathe use of multi-reservations. With the traditional

20

70

120

170

220

0 10 20 30 40 50Level of Heterogenity

Mea

n D

elay

(ms)

AODV +IEEE802.11 MAC DSR +IEEE802.11 MAC

DSR +Cross Layer Framework AODV +Cross Layer Framework)

0

2

4

6

8

10

12

14

16

18

10 20 25 30 40 50Level of Heterogeneity (%)

% In

crea

se in

Nor

mal

ized

Rou

ting

Load

(by

tes)

DSR AODV

Fig. 16. End-to-end delay and routing overhead with ourframework. (a) End-to-end delay. (b) Routing overhead withour cross-layer framework in the heterogeneous network.

0

1

2

3

4

5

6

7

10 20 30 40 50

Level of Heterogenity (%)

% F

ailu

re fo

r Mul

ti-

Res

erva

tion

Sche

me

with MAC Enhancements Cross-Layer Framework

40

45

50

55

60

65

70

75

80

0 5 10 15

Speed (meters per second)

Size

of u

pdat

e m

essa

ges

(byt

es)

20

1-Hop 2 Hops 3 Hops

60

80

100

120

nod

es c

over

ed

Two Hop Coverage


schemes, route failures and the consequent routediscovery attempts occur with greater frequency,and during these periods data packets simply waitin the source queue. At higher levels of heterogene-ity, the overall delays experienced at the MAC layerincreases, due to increased collisions and retrans-missions caused by asymmetry (Fig. 16b). Howeverwe still find that the mean delay with our cross-layerframework is marginally larger than that with tradi-tional protocols. We believe that this slight increasein delay is acceptable, considering that with ourframework we observe significant gains in the over-all throughput efficiency and packet delivery ratio.

0

20

40

1 3

Number of Transmissions heard (T)

% In

terfe

ring

2 4 5

Fig. 17. System parameters with our scheme. (a) Packet failureswith the multi-reservation scheme when N = 3. (b) Size of updateHello messages at different levels of mobility. (c) Choice ofparameter T.

5.3. Choice of system parameters

In this subsection we study the sensitivity of oursimulation results to the system parameters thatwere used. We have earlier presented intuitive justi-fications for our selected values of the systemparameters; in the following, we provide exper-

imental results that corroborate our previousdiscussions.

Choice of T: T is a parameter associated with theIntelligent Broadcasting scheme; if a node overhearsbroadcasts from at least T neighbors, it quells itsown. As depicted in Fig. 17c, when T is 1 or 2, nodesare over-aggressive in quelling their broadcasts. Asa result, the coverage in terms of reaching all thepotential interfering high power nodes is small.


For values of T strictly greater than 3, nodes end upperforming more re-broadcasts than needed, thusrendering the additional coverage benefits negligi-ble. Thus, we set T = 3.

Choice of N: To recap, N is the number of pack-ets (intended for sequential transmissions), forwhich a node can simultaneously reserve the chan-nel with a single RTS/CTS/BW_RES instantiationusing the multi-reservation scheme. In Fig. 17a weplot the percentage of times when a packet failswhile using the multi-reservation scheme with vari-ous levels of heterogenity, when the value of N isfixed to 3. We plot, (a) the performance using theMAC layer enhancements with our Poisson trafficagent at the MAC layer and (b) the performanceof our cross-layer framework. The percentage offailures observed with the latter is higher, due tothe additional overhead of routing. Failures occureither due mobility or because our schemes areunable to reach all potential interferer high powernodes (unavailability of paths). The latter effect alsoincreases with the level of heterogeneity (increasedasymmetry) as borne out by the figure. Failure ratesare nominal with N = 3 (between 3% and 6%). Athigher values of N we observe that failure ratesincrease considerably (we do not report results heredue to space constraints). This is because a highernumber of back-offs15 occur for larger values of N.In addition, larger values of N also lead to the dom-inance of the channel by certain nodes therebyincreasing the unfairness in channel access. Thus,in our simulations we use a small value of N, specif-ically, 2 for high power nodes and 3 for low powernodes. We provide a larger value of N for low powernodes in order to improve fairness for these nodes intheir channel access procedure.

Choice of n: The choice of n governs the ability ofnodes to find reverse routes for bridging a unidirec-tional link. As mentioned earlier, n refers to the dis-tance in terms of hop count, up to which a nodecollects and disseminates local neighborhoodinformation. We know from our earlier intuitivediscussions that n should be at least 2 in order thata low power node can identify the high power nodesthat are beyond its own transmission range. Weobserve that the size of the n-hop neighborhood isrelatively unaffected by node mobility, as we useconstant speeds in our model. Fig. 17b depicts thatthe size of the route update packets increases withspeed and with n. Since with higher mobility the

15 A transmission failure causes a node to back-off.

localized neighborhood of a node is likely to changemore often, the size of these update Hello messagesincreases with mobility. A higher value of n causesan increase in the size of the update messages sincenodes have larger localized graphs. We remark thatan update message contains only the changes in thelocalized graph incurred since the previous updatemessage. Since, the average increase in the size ofthe update messages with an increase in n from 2to 3 is nominal (increases by about 10 bytes16) weuse n = 3. We find that this value improves the per-formance, since a large number of unidirectionallinks are now bridged and a sufficiently high numberof interfering high power nodes are prevented frominitiating transmissions when a low power commu-nication is in progress.

6. Conclusions and future work

In this paper our key contribution is the develop-ment of a unified framework that offers a couplingbetween the MAC and the routing layers to dealwith link level asymmetries in power heterogeneousad hoc networks. While there had been no priorsolutions that handle asymmetry effectively at theMAC layer, previous unidirectional routing schemeshad ignored the MAC layer dependencies. In ourframework the MAC layer solicits assistance fromthe routing layer to identify link asymmetry. Lowpower nodes then route MAC layer controlframes to high power nodes that are beyond theirtransmission range, to inhibit them from perform-ing transmissions while they are in the process ofcommunicating with other nodes. At the same time,the framework also allows for the identification andusage of unidirectional links at the routing layer.This in turn leads to shorter routes and conse-quently to improved performance. We also consid-ered two techniques that are exclusively deployedat the MAC layer based on (a) the use of an intelli-gent broadcast scheme to quell unnecessary broad-casts, and (b) reserving the bandwidth for multipledata frames with a single RTS/CTS exchange. Wefind that while these schemes can also provideimprovements in performance, our cross-layerapproach does significantly better. We study theperformance exclusively at the MAC layer and atthe higher layers. At the transport layer, ourcross-layer framework can improve throughput at

16 Note that the IEEE 802.11 data frame is typically of the orderof 2 K Bytes.


the low power nodes by up to 25%, alleviating theunfairness introduced with the legacy IEEE 802.11MAC. We also show a significant reduction (by20%) in the total number of false link failurescaused in the network due to interference fromneighboring nodes. As a result of this comprehen-sive simulation-based analysis, our integratedMAC/Routing layer framework is shown to offera simple yet viable and effective solution for han-dling asymmetry in power heterogeneous ad hocnetworks.

In this work, we propose the use of traditionalon-demand routing protocols (DSR and AODV)atop our integrated framework. While this is apreferable approach for ensuring backward compat-ibility with possible previous implementations ofthese protocols, one might also explore future opti-mizations to network wide routing in power hetero-geneous networks. In addition, in this work we havesimulated fixed power levels and therefore coarse-grained power heterogeneity. Our schemes need tobe examined when nodes deploy fine-tunable powerlevels, i.e. in the presence of power control. There isa trade-off between exploiting spatial re-use withpower control and dealing with the consequent linklevel asymmetry. We are planning to investigatethese aspects in our future efforts.

References

[1] IEEE STD 802.11 1999. Wireless LAN Medium AccessControl (MAC) and Physical Layer (PHY) Specifications,1999.

[2] Y. Afek, E. Gafni, Distributed algorithms for unidirectionalnetworks, SIAM Journal of Computing 23 (6) (1994) 1152–1178.

[3] L. Bao, J.J. Garcia-Luna-Aceves, Link-state routing innetworks with unidirectional links, in: IEEE InternationalConference on Computer Communications and Networks(IC3N), 1999.

[4] L. Bao, J.J. Garcia-Luna-Aceves, Channel access schedul-ing in ad hoc networks with unidirectional links, in:ACM Worskshop on the Discrete Algorithms and Methodsfor Mobile Computing and Communication (DIAL-M),2001.

[5] V. Bhargavan, A. Demers, S. Shenker, L. Zhang, Macaw: amedia access protocol for wireless lans, in: ACM SIG-COMM, 1994.

[6] D.M. Blough, M. Leoncini, G. Resta, P. Santi, The litK-neigh protocol for symmetric topology control in ad hocnetworks, in: ACM MOBIHOC, 2003.

[7] M. Burkhart, P.v. Rickenbach, R. Wattenhofer, A. Zollin-ger, Does topology control reduce interference? in: ACMMOBIHOC, 2004.

[8] G. Calinescu, Computing 2-hop neighborhoods in ad hocwireless networks, in: ADHOC-NOW, 2003.

[9] E. Duros, W. Dabbous, H. Izumiyama, N. Fujii, Y. Zhang,A link-layer tunneling mechanism for unidirectional links,RFC 3077, 2001.

[10] Z.J. Haas, M.R. Pearlman, The zone routing protocol (ZRP)for ad hoc networks, Internet draft-Mobile Ad hocnetworking (MANET) Working Group of the InternetEngineering Task Force (IETF), March 2001.

[11] Eun-Sun Jung, Nitin H. Vaidya, A power control macprotocol for ad hoc networks, in: ACM MOBICOM, 2002,pp. 36–47.

[12] D. Kim, C.K. Toh, Y. Choi, GAHA and GAPA: two link-level approaches for supporting link asymmetry in mobile adhoc networks, IEICE Transaction on CommunicationsE-86B (4) (2003) 1297–1306.

[13] T.J. Kwon, M. Gerla, Clustering with power control, in:IEEE MILCOM, vol. 2, 1997, pp. 919–923.

[14] E.L. Li, J.Y. Halpern, P. Bahl, Y.M. Wang, R. Wattenhofer,Analysis of a cone-based distributed topology controlalgorithm for wireless multi-hop networks, in: ACM PODC,2001.

[15] N. Li, J.C. Hou, L. Sha, Design and analysis of an mst-basedtopology control algorithm, in: IEEE INFOCOM, 2003.

[16] X.Y. Li, W. Song, W. Wang, A unified energy-efficienttopology for unicast and broadcast, in: ACM MOBICOM,2005.

[17] X.Y. Li, P.J. Wan, Y. Wang, O. Frieder, Sparse powerefficient topology for wireless networks, in: IEEE HICSS,2002.

[18] W. Liu, Y. Zhang, W. Lou, Y. Fang, DELAR: Device/energy/load aware relaying in heterogenous wireless ad hocnetworks, in: Military Communications Conference (MIL-COM), vol. 3, 2004, pp. 1303–1309.

[19] Mahesh K. Marina, Samir Ranjan Das, Routing perfor-mance in the presence of unidirectional links in multihopwireless networks, in: ACM MOBIHOC, 2002.

[20] J.P. Monks, V. Bharghavan, W.-M.W. Hwu, A powercontrolled multiple access protocol for wireless packetnetworks, in: IEEE INFOCOM, 2001, pp. 219–228.

[21] Alaa Muqattash, Marwan Krunz, Power controlled dualchannel (pcdc) medium access protocol for wireless ad hocnetworks, in: IEEE INFOCOM, 2003.

[22] Alaa Muqattash, Marwan Krunz, A single-channel solutionfor transmission power control in wireless ad hoc networks,in: ACM MOBIHOC, 2004, pp. 210–221.

[23] S. Narayanaswamy, V. Kawadia, R.S. Sreenivas,P.R. Kumar, Power control in ad hoc networks: theory,architecture, algorithm and implementation of the compowprotocol, in: European Wireless Conference, 2002, pp. 156–162.

[24] S. Nesargi, R. Prakash, A tunneling approach to routingwith unidirectional links in mobile ad hoc networks, in:IEEE International Conference on Computer Communica-tions and Networks (IC3N), 2000, pp. 522–527.

[25] S. Ni, Y. Tseng, Y. Chen, J. Sheu, The broadcast stormproblem in a mobile ad hoc network, in: ACM MOBICOM,1999.

[26] N. Poojary, S.V. Krishnamurthy, S. Dao, Medium accesscontrol in a network of ad hoc nodes with heterogeneoustransmit power capabilities, in: IEEE ICC, 2001.

[27] R. Prakash, A routing algorithm for wireless ad hocnetworks with unidirectional links, ACM/Kluwer WirelessNetworks 7 (6) (2001) 617–625.


[28] Ram Ramanathan, Regina Hain, Topology control ofmultihop wireless networks using transmit power adjust-ment, in: IEEE INFOCOM, 2000, pp. 404–413.

[29] Venugopalan Ramasubramanian, Ranveer Chandra,Daniel Mosse, Providing a bidirectional abstraction forunidirectional adhoc networks, in: IEEE INFOCOM,2002.

[30] E. Royer, C-K. Toh, A review of current routing protocolsfor ad hoc wireless networks, IEEE Personal Communica-tions Magazine 5 (1999).

[31] P. Sinha, S.V. Krishnamurthy, S. Dao, Scalable unidirec-tional routing using zrp extensions for wireless networks, in:IEEE WCNC, 2000.

[32] W. Song, Y. Wang, X. Li, Localized algorithms for energyefficient topology in wireless ad hoc networks, in: ACMMOBIHOC, 2004.

[33] S.L. Wu, Y.C. Tseng, J.P. Sheu, Intelligent medium accessfor mobile ad hoc networks with busy tones and powercontrol, IEEE Journal on Selected Areas in Communications(JSAC) 18 (9) (2000) 1647–1657.

[34] Jungkeun Yoon, Mingyan Liu, Brian Noble, Randomwaypoint considered harmful, in: IEEE INFOCOM, 2003.

Vasudev Shah received his B.S. in Elec-trical Engineering at the University ofBombay, India in 2000, M.Sc. in Com-puter Science at the University of Cali-fornia, Riverside in 2004 and he is aPh.D. Candidate at the University ofCalifornia-Riverside. Currently he is aChief Technology Officer in First Net-work Solutions, a Los Angeles basedcompany that specializes in the designand implementation of WAN solutions,

Voice and Video over IP solutions for Medium and Large busi-nesses. His research interests are the design of medium access
control (MAC) and routing protocols, and power controlmechanisms in wireless ad hoc networks.
Ece Gelal received her B.Sc. in Tele-communications from Sabanci Univer-sity, Turkey and her M.Sc. in ComputerScience and Engineering from the Uni-versity of California, Riverside in 2004and 2006, respectively. Currently she is aPh.D. candidate in Computer Scienceand Engineering in the University ofCalifornia, Riverside. Her researchinterests are cross-layer design approa-ches for mobile ad hoc networks

(MANETs), power control, topology control in ad hoc networks,and effective use of smart antennas in ad hoc networks via
properly designed MAC and routing protocols.
Srikanth V. Krishnamurthy received hisPh.D. degree in electrical and computerengineering from the University of Cali-fornia at San Diego in 1997. From 1998to 2000, he was a Research Staff Scientistat the Information Sciences Laboratory,HRL Laboratories, LLC, Malibu, CA.Currently, he is an Associate Professorof Computer Science at the University ofCalifornia, Riverside. His researchinterests span CDMA and TDMA tech-

nologies, medium access control protocols for satellite andwireless networks, routing and multicasting in wireless networks,
power control, the use of smart antennas and security in wirelessnetworks. He has been a PI or a project lead on projects fromvarious DARPA programs including the Fault Tolerant Net-works program, the Next Generation Internet program and theSmall Unit Operations program. He is the recipient of the NSFCAREER Award from ANI in 2003. He has also co-edited thebook ‘‘Ad Hoc Networks: Technologies and Protocols’’ pub-lished by Springer Verlag in 2005. He has served on the programcommittees of INFOCOM, MOBIHOC and ICC and is theassociate editor-in-chief for ACM MC2R.

handling asymmetry in power heterogeneous ad hoc networkskrish/comnet-power.pdf · heterogeneous...

Documents