Ad Hoc Networks 8 (2010) 755–766

Contents lists available at ScienceDirect

Ad Hoc Networks

journal homepage: www.elsevier .com/locate /adhoc

Energy-efficient proactive routing in MANET: Energy metrics accuracy

Thomas Kunz a,*, Rana Alhalimi b

a Systems and Computer Engineering, Carleton University, 1125 Colonel By Drive, Ottawa, Ont., Canada K1S 5B6b School of Computer Science, Carleton University, 1125 Colonel By Drive, Ottawa, Ont., Canada K1S 5B6

a r t i c l e i n f o a b s t r a c t

Article history:Received 1 May 2009Received in revised form 1 February 2010Accepted 8 February 2010Available online 12 February 2010

Keywords:MANETsState information accuracyEnergy levelQuality of Service (QoS)

1570-8705/$ - see front matter � 2010 Elsevier B.Vdoi:10.1016/j.adhoc.2010.02.004

* Corresponding author. Tel.: +1 613 520 3573; faE-mail addresses: tkunz@sce.carleton.ca (T.

carleton.ca (R. Alhalimi).

To support energy-efficient routing, accurate state information about energy levels shouldbe available. But due to bandwidth constraints, communication costs, high loss rate and thedynamic topology of MANETs, collecting and maintaining up-to-date state information is avery complex task. In this work, we use Optimized Link State Routing (OLSR) as the under-lying routing protocol and explore the accuracy of state information under different trafficrates. We are focusing on energy level as QoS metric, which has been used for routing deci-sions in many energy-efficient routing protocol proposals. First, we show that the accuracyof the available nodal energy level does impact the performance of energy-efficient varia-tions of OLSR. If nodes learn other nodes’ energy level through protocol messages, fewerpackets tend to get delivered in an energy-constrained network, in particular under hightraffic loads or in mobile networks. We analyzed the accuracy of the reported energy levelsfor the static scenarios and found that the propagated values are highly inaccurate, in par-ticular under high traffic rates. Tuning the OLSR protocol parameters has no noticeableimpact on accuracy levels. We then propose two additional techniques to increase accura-cies and compare the different techniques against each other and against the basic OLSRprotocol. One of the techniques, which we call smart prediction, achieves highly accurateperceived energy levels under all traffic loads. We finally show that the proposed smartprediction technique also works well for mobile networks and more heterogeneous wire-less interfaces.

� 2010 Elsevier B.V. All rights reserved.

1. Introduction

Optimized Link State Routing (OLSR) is a routing proto-col used for Mobile Ad-Hoc Networks (MANET) [1]. It is abest-effort proactive protocol. Proactive protocols are char-acterized by all nodes maintaining routes to all destina-tions at all times through the periodic exchange ofprotocol messages. This gives them the advantage of hav-ing pre-computed routes available when needed and topropagate topology changes in bulk updates to manynodes. OLSR performs hop-by-hop routing, where eachnode uses its most recent topology information for routing.

. All rights reserved.

x: +1 613 520 5727.Kunz), ralhali2@scs.

OLSR is highly focused on reducing the protocol over-head. As a result, information about QoS-related state isnot propagated throughout the network. But with the ris-ing popularity of multimedia applications and the poten-tial commercial usage of MANETs, QoS support in ad-hocnetworks has become a very critical issue and a range ofQoS signaling and routing protocols have been proposed.

To support QoS routing, state information such as en-ergy level, bandwidth or queue length should be availablewhen making routing decisions at a node. But because thequality of wireless links changes quite frequently due tomobility and changes in surroundings, coupled with thelimited wireless bandwidth, collecting and updating suchknowledge is a non-trivial task.

Since OLSR is based on a periodic exchange of messages,QoS-related state information might not be up-to-date at

756 T. Kunz, R. Alhalimi / Ad Hoc Networks 8 (2010) 755–766

any instance of time. Therefore, nodes might have inaccu-rate information about other nodes in the network, whichmight have a negative impact on the performance of thenetwork. The motivation of this research is to quantifythe accuracy of the QoS-related state information in ad-hoc networks under different conditions and if possible,devise techniques to reduce inaccuracies.

State information represents a QoS-related state. Itcould be node attributes such as energy level and queuelength, or link attributes such as bandwidth, delay, or errorrate. In this study, we are interested in node attributes, andspecifically energy level. Due to energy constraints, a num-ber of energy-aware routing protocols have been proposed[2–4]. Our work investigates the possibility of makingOLSR an efficient energy-aware routing protocol by evalu-ating how accurately individual nodes learn about the cur-rent energy levels for other nodes when piggybacking thisinformation onto OLSR control messages.

State information accuracy specifies how accurate is theavailable QoS-related state information relative to their ac-tual values. Throughout this paper, we refer to the latter asthe actual value and the former as the perceived value. Themain contributions of this work are:

� Demonstrating the impact of inaccurate QoS informa-tion on energy-efficient modifications of OLSR.

� Quantification of the inaccuracy of energy level informa-tion in ad-hoc networks under different traffic loads,using OLSR as the underlying routing protocol.

� Studying the impact of tuning the OLSR protocol param-eters on the inaccuracy level of the energy level metric.

� Suggesting and developing techniques to reduceinaccuracies.

� Evaluating the performance of the proposed techniquesand comparing them to the basic OLSR protocolperformance.

The paper is organized as follows: a review of relatedwork in the area of QoS state information accuracy is pre-sented in Section 2. Section 3 summarizes the core func-tionality of the Optimized Link State Routing protocol(OLSR) and describes our changes to the protocol to prop-agate QoS state information. Section 4 motivates our work,showing that simply propagating a node’s residual energylevel through control messages results in lower protocolperformance for energy-efficient variants of OLSR. Section 5analyzes, using extensive simulations, the levels andcauses of inaccurate QoS state information. This sectionalso analyzes the effect of varying various protocol param-eters on the inaccuracy level. Section 6 discusses two tech-niques for improving the overall energy inaccuracy leveland evaluates their performance through simulations. Sec-tion 7 expands this discussion to mobile scenarios and het-erogeneous radio interfaces. The final section concludesour work and suggests possible future work.

2. Related work

OLSR is a well-known routing protocol for ad-hoc net-works. It has been broadly examined [5–8], implemented

and deployed [9,10]. Performance measurements in a realtest-bed are reported in [6], where the authors concludethat OLSR suffers from high variability of performancedepending on how far apart are the nodes, and from unfair-ness depending on the topology and on the nature of thetraffic. It suggests that QoS features could complementthe performance of the OLSR protocol.

Ge et al. [11] develops a QoS version of the OLSR proto-col, based on link bandwidth as QoS metric. This QoS OLSRprotocol attempts to find paths with maximum bottleneckbandwidth. In order to support QoS (provide optimal band-width path), changes in the link bandwidth must be prop-agated for the correct computation of the best bandwidthroute. The authors evaluate the performance of this QoSOLSR model under different bandwidth change thresholdvalues and compare it to the basic OLSR protocol perfor-mance. These threshold values define a tradeoff betweenthe accuracy of link-bandwidth information and the addi-tional overhead the routing protocol introduces. Threethreshold values (20%, 40% and 80%) are used. The resultsshow that amongst the proposed QoS OLSR algorithms,20% QoS OLSR calculates the routes that are closest to theoptimal routes compared to the 40% and 80% QoS OLSR.This is due to the fact that 20% QoS OLSR updates the band-width condition most frequently and consequently getsthe most accurate bandwidth information. Ge et al. [11]demonstrates that OLSR has a potential for QoS routing.It also shows that the availability of more accurate stateinformation throughout the network, via more frequentupdates, improves the performance of QoS routing. How-ever, their work does not investigate quantitatively the le-vel of accuracy of the QoS metric (link bandwidth).

Clausen et al. [5] investigates the impact of extendingtopology knowledge on the OLSR protocol performance.In an OLSR network, nodes have partial topology knowl-edge in which only a subset of the links are known to thenode to reduce the protocol overhead. Increasing the par-tial topology information provides a more robust and accu-rate topology view. It is achieved by increasing the numberof links advertised and number of nodes advertising links.In OLSR this can be done by varying two protocol parame-ters (named MPR-coverage and TC-redundancy). In orderto determine the effect of advertising redundant and moreaccurate topology information on the performance of theOLSR routing protocol, Clausen et al. [5] studies the impactof increasing the MPR-coverage parameter. Their resultsshow higher packet delivery rates under moderate nodemobility when increasing the redundancy of topologicalinformation and retransmissions provided by a higherMPR-coverage.

The research in [12] expands the work done in [5] andentirely focuses on understanding the tradeoffs of increas-ing accurate topology knowledge. It investigates the im-pact of tuning the MPR-coverage and TC-redundancyparameters on the OLSR performance. It shows that deliv-ery rates are not affected by the overhead resulting fromadvertising redundant information. Both [5,12] focus onhaving more accurate information at the topology (net-work) level and how it affects the routing protocol perfor-mance. In other words, they study the effect of tuning theOLSR protocol parameters on accuracy in terms of network

T. Kunz, R. Alhalimi / Ad Hoc Networks 8 (2010) 755–766 757

status (existing nodes and links) and not in terms of stateinformation available at nodes and links.

Various researchers [13–15] discuss energy-efficientvariants of OLSR. Ghanem et al. [13] suggest modificationsto the MPR selection mechanism, choosing MPRs based ontheir residual energy levels, rather than on their coverageof 2-hop neighbors, as suggested by the original protocolRFC [1]. Guo and Malakooti [14] modify the path determi-nation algorithm, selecting paths based on the residual en-ergy level of intermediate nodes, selecting routes thatavoid nodes with low residual energy. Benslimane et al.[15] combine both MPR selection and path determinationto again select paths with maximum bottleneck residualenergy level. All papers publish simulation results thatshow that, for certain scenarios, taking energy constraintsinto consideration can increase nodal lifetime and networkperformance, yet are silent on how to collect informationabout residual energy levels accurately. We evaluated allthese variants [7] through simulations, assuming thatnodes have omniscient knowledge about a node’s residualenergy. In this case, using the original MPR selection crite-ria (which aims to minimize the number of MPRs), to-gether with a path determination algorithm that avoidsnodes with low residual energy provides the best perfor-mance. For many scenarios, in particular as node mobilityincreases, changing the MPR selection criteria (in additionto the new path determination algorithm) is beneficial aswell. More recently, Mahfoudh and Minet [8] proposed avariant of OLSR called EOLSR, where MPR selection andpath calculation is determined by both a node’s residualenergy level and its number of neighbors. The key insighthere is that sending data to a node also forces all its neigh-bors to consume energy in overhearing the data packet.The simulation results reported in [8] show that combiningboth the new path calculation with the modified MPRselection yields the best performance. The paper suggeststhat a node’s residual energy level is propagated byextending the protocol control messages, but does not dis-cuss how accurate this information is.

The most relevant body of work to the problem of QoSrouting in the presence of inaccurate information is a setof papers aimed at exploring state-aggregation issues andtheir impact on routing performance in large networks.They emphasize on developing good aggregation tech-niques that minimize inaccuracy in network state informa-tion, while allowing substantial reductions in the amountof state data. Han and Venkatasubramanian [16] addressthe information collection problem for QoS-based servicesin mobile environments. Specifically, they propose a familyof information collection policies that vary in the granular-ity at which system state information is represented andmaintained. The authors evaluate the impact of informa-tion collection algorithms on the performance of QoS-based resource provisioning. The work in [16] proposestwo approaches to collecting location information for mo-bile applications. Fine-grained approaches maintain cur-rent location of each individual mobile client, whilecoarse-grained collection captures information at anaggregate level of multiple clients. Han and Venkatasubra-manian [16] conclude that coarse-grained mobility infor-mation is sufficient for effective resource provisioning,

whereas fine-grained mobility information introduces avery high overhead. The work however does not evaluatethe impact of fine-grained mobility information on re-source provisioning and how it compares to coarse-grainedmobility information. Since we are dealing with relativelysmaller networks, our work investigates the impact of col-lecting fine-grained data on the accuracy of state informa-tion as an upper bound on achievable accuracy.

A few researchers [17–19] investigate the impact ofinaccuracies, in the available network state and metricinformation, on the path selection process for flows whichrequire QoS guarantees. Shaikh et al. [19] in particularevaluate the impact of inaccurate state information onthe performance and overheads of QoS routing by evaluat-ing periodic and triggered updates. The paper uses connec-tion blocking as performance measure. Connectionblocking defines the percentage of times a connection re-quest from a source to a destination fails. The authors drawa distinction between routing failures and setup failures.Routing failures occur when the source cannot compute afeasible path for the new connection. In contrast, setupfailures occur when the source selects a seemingly feasiblepath that ultimately cannot support the new connection.With a periodic update policy, larger periods substantiallyincrease connection blocking, ultimately outweighing thebenefits of QoS routing. The results show that a periodicupdate policy alone cannot meet the dual goals of lowblocking probability and low overhead in realistic net-works. In contrast, experiments with triggered updatesshow that coarse-grained triggers do not have a significantimpact on the overall blocking probability, although largertriggers shift the type of blocking from routing failures tomore expensive setup failures.

In summary, none of the prior works fully address theproblem of inaccurate state information. Some work iscompletely silent on the problem, simply assuming thatQoS-related state information needs to be collected andpropagated throughout the network. Others acknowledgethat the collection and dissemination of QoS state inevita-bly will cause inaccurate information and show qualita-tively a tradeoff between state information accuracy andprotocol performance. In the work reported here, we focusexplicitly on the inaccuracy of state information, more spe-cifically the residual energy level of nodes as used by en-ergy-efficient MANET routing protocols. Using OSLR asbase routing protocol, we show that inaccurate informa-tion does reduce protocol performance, quantify theamount of inaccuracy, analyze the reasons, and suggestalternatives to increase state information accuracy.

3. Description of OLSR and modifications

The IETF Working Group introduced the Optimized LinkState Routing protocol (OLSR) for mobile ad-hoc networks[1]. The protocol is an optimization of the pure link statealgorithm. The key concept used in the protocol is that ofmultipoint relays (MPRs). Each node selects a set of itsneighbor nodes as MPRs. Only nodes selected as MPRsare responsible for forwarding control traffic, intendedfor diffusion into the entire network. MPRs provide an

758 T. Kunz, R. Alhalimi / Ad Hoc Networks 8 (2010) 755–766

efficient mechanism for flooding control traffic by reducingthe number of (re-)transmissions required.

Nodes selected as MPRs also have a special responsibil-ity when declaring link-state information in the network.Indeed, the only requirement for OLSR to provide shortestpath routes to all destinations is that MPR nodes declarelink-state information for their MPR selectors.

Nodes which have been selected as multipoint relays bysome neighbor node(s) announce this information period-ically in their control messages. Thereby a node announcesto the network that it has reachability to the nodes whichhave selected it as an MPR. In route calculation, the MPRsare used to form the route from a given node to any desti-nation in the network. Furthermore, the protocol uses theMPRs to facilitate efficient flooding of control messagesin the network.

Due to its proactive nature, OLSR works with a periodicexchange of messages. The key messages are Hello and TCmessages. Hello messages are periodically exchanged toinform nodes about their neighbors and their neighbors’neighbors and are 1-hop broadcast messages. The 2-hopneighborhood information is then used locally by eachnode to determine MPRs. In contrast, TC messages areflooded through the network to inform all nodes aboutthe (partial) network topology. At a minimum, TC mes-sages contain information about MPRs and their MPRselectors.

To control the protocol overhead, OLSR defines a fewparameters. The Hello-interval parameter (default: 2 s)represents the frequency of generating a Hello message.Increasing the frequency of generating Hello messagesleads to more frequent updates about the neighborhoodand hence a more accurate view of the network.

The TC-interval parameter (default: 5 s) represents thefrequency of generating a TC message. TC messages areone of the major sources of overhead in MANETS, as theyare flooded throughout the network, but they facilitatethe topology discovery process. Since nodes learn aboutthe whole topology by exchanging TC messages, the morefrequently nodes generate TC messages, the more recentthe knowledge nodes have about the topology.

The MPR-coverage parameter (default: 1) allows a nodeto select redundant MPRs. The criterion for selecting MPRsis that all 2-hop neighbors must be reachable through atleast one MPR node. Nodes should select their MPR set tobe as small as possible in order to reduce protocol over-head. Redundancy of the MPR set affects the overheadthrough affecting the amount of links being advertised,since a node will be selected by more neighbor nodes asan MPR, the amount of nodes advertising links, since morenodes will be selected as MPRs, and the efficiency of theMPR flooding mechanism. On the other hand, redundancyin the MPR set ensures that reachability for a node isadvertised by more nodes.

The TC-redundancy parameter (default: 0) specifies, forthe local node, the amount of information that may be in-cluded in the TC message. A TC-redundancy of 0 specifiesthat the advertised link set of the node is limited to linksto its MPR selectors. A TC-redundancy of 1 specifies thatthe advertised link set of the node is the union of links toits MPR selectors and to other MPRs. A TC-redundancy of

2 specifies that the advertised link set of the node is the fullneighbor link set. The TC-redundancy parameter affectsthe overhead through affecting the amount of links beingadvertised as well as the amount of nodes advertisinglinks.

In order to quantify the accuracy of state information,the QoS-related state needs to be propagated throughoutthe network. There are two ways in which QoS-relatedstate can be propagated throughout the network. Eitherwe define a new message type to carry the QoS-relatedstate information, or we include it in the OLSR protocolmessages (Hello and TC messages) to be available to othernodes in the network. With the first approach, a new mes-sage type has to be defined and exchanged. This will incura potentially large overhead in the network since moremessages will be exchanged. In addition, these messageswill include a lot of redundant information, compared tothe existing control messages. We therefore choose to ex-tend the existing control messages by including the QoS-related state information in the OLSR protocol messages,as suggests in [8] as well.

Through the exchange of OLSR control messages, eachnode accumulates information about the network. Thisinformation is stored according to the OLSR specifications.To store the QoS-related state associated with a node, anew field was added to the neighborhood information baseand to the topology information base maintained by theprotocol.

Extended Hello messages are broadcast to all one-hopneighbors. They contain not only a list of addresses ofneighbors, but also the most recent QoS-related state asso-ciated with those neighbors from the sender node’s per-spective. In addition to that, the message also containsthe QoS-related state of the sender node itself at the timethe message is generated. The other fields are loadedaccording to the OLSR specifications.

Extended TC messages are broadcast and retransmittedby the MPRs in order to diffuse topology information intothe entire network. TC messages contain not only a list ofaddresses of a node’s MPR selectors, but also the QoS-re-lated state associated with those nodes from the originatornode’s perspective. In addition to that, the message alsocontains the QoS-related state of the originator node atthe time the message is generated.

One consequence of these modifications is that, for a gi-ven node X, a single node may have multiple differentresidual energy levels stored in various internal databases.Also, as a result of message delays and message losses, anode may learn old information about node X from other,intermediate nodes, at a later point in time. To disambigu-ate between these entries and to determine the most re-cent values, we associate a timestamp with each datapoint and modify the control messages and local reposito-ries accordingly. As energy level is a monotonicallydecreasing value, a simpler solution would be to alwaysuse the smallest known value (which would be the mostrecent). However, this does not work if we were to exploreother, non-monotonic, QoS metrics such as queue lengths(for load-balanced routing) or link bandwidth. No clocksynchronization is required if these timestamps are onlyused to compare values originating from the same node

T. Kunz, R. Alhalimi / Ad Hoc Networks 8 (2010) 755–766 759

(to determine the most recent value). However, we alsoutilize these timestamps to analyze delays and ‘‘knowledgeage”, requiring the network nodes to be time-synchro-nized. The results reported in this work are based on theNS2 simulator, so this is given (i.e., all nodes referencethe same global clock). For real deployments, as our analy-sis deals with delays in the order of seconds, no particu-larly tight clock synchronization (i.e., at the level ofmicroseconds) is required and achievable with one of themany proposed clock synchronization algorithms.

4. Motivating example

In past research, we explored how to modify OLSR totake energy constrains into consideration to increase anode’s lifetime and provide improved routing protocol per-formance [7]. We report here results for two sets of modi-fications to OLSR: we modified the MPR selection, and wemodified the path determination algorithm. MPRs are asubset of the 1-hop neighbors that provide access to all2-hop neighbors of a node. As most routing is done viaMPR, we iteratively add 1-hop neighbors with maximalresidual energy level to the MPR set until all 2-hop neigh-bors are covered. More formally, when a node N selects itsMPRs:

– N puts in the set Uncovered all its strict two-hopneighbors.

– N sorts its 1-hop neighbors by decreasing residualenergy level, let N1 be this ordered set.

– Repeat:� Node N selects the first node in N1 (the node with the

highest residual energy level) and removes it fromN1, let that node be M.

� If M covers any nodes in Uncovered, M becomes anMPR, remove the nodes covered by M fromUncovered.

– until Uncovered becomes empty.

This algorithm will increase the number of MPRs se-lected and therefore the protocol overhead to flood TCmessages. As discussed in [8], where MPR selection is alsomodified to take residual energy into account, the increasemay be substantial (log(n), where n is the number of neigh-bors of a node). However, as routes ultimately are buildfrom MPRs (except potentially the first and last node), thiswill avoid nodes with low residual energy levels, unlessthey are the source or destination. In addition, as reportedin [7], such a modification, in conjunction with a change inthe route selection, will result in better protocol perfor-mance when energy is a scarce resource.

The path determination algorithm in the original OLSRprotocol is essentially a Dijkstra shortest-path algorithm.Nodes learn a partial network topology, over which theythen perform shortest-path routing to populate the routingtable. Our modification changes the weight associated witheach link. Rather than assigning each link the same con-stant weight of 1, we assign it the reciprocal value of thesending node’s residual energy level. Again, this willpenalize routes that traverse nodes with low residual

energy level. In addition, using the reciprocal value hasthe advantage that the algorithm does not need ‘‘artificial”thresholds to determine when a node’s residual energylevel is low and the link should therefore be assigned ahigher weight. As a node’s energy level depletes, the pathdetermination algorithm will be increasingly less likely touse such nodes, as the associated costs are increasingrapidly.

As discussed previously, we found that the best overallperformance across different traffic loads and mobility sce-narios is obtained by using the new path determinationalgorithm, combined with the original MPR selection crite-ria. For many scenarios, in particular as node mobility in-creases, changing the MPR selection criteria (in additionto the new path determination algorithm) is beneficial aswell. So our results here cover two protocol variants: Mod-ified Routing refers to the version that uses the original MPRselection criteria, but uses the new path determinationalgorithm, whereas Modified MPR/Routing combines boththe new MPR selection and the new path determinationalgorithm. We use two different versions of each protocol,the ideal version (where a node has access to the actual,current residual energy level of remaining nodes whenselecting MPRs and determining paths) and the realisticversion, in which a node needs to rely on the residual en-ergy levels it learned through protocol messages as de-scribed above.

We performed extensive simulation studies for thesetwo protocol variants and the two ways of learning anode’s residual energy level, for different mobility ratesand network traffic levels. The details of the common sim-ulation parameters are similar to the ones described later,and not reported here for the sake of brevity. The only ma-jor difference to the latter information is that we set theinitial energy level of each node to a small value (15 J) togenerate scenarios where nodes involved in forwardingdata packets deplete their energy and die. We also evalu-ated two different levels of mobility. In the static scenarios,nodes are randomly placed into the network, but do notmove during the simulation. In the mobile scenarios, nodesmove around the simulation area based on the RWP mobil-ity model, with a maximum speed of 2 m/s and a pausetime of 10 s. The following two figures summarize ourfindings; each data point represents the average of at least10 simulation runs for different initial node placementsand (where applicable) node movements.

RFC 2501 [20] describes performance metrics for theevaluation of routing protocols and has been widelyadopted. We focus on only one parameter here, and thatis the network’s ability to deliver data packets. This is tra-ditionally expressed as Packet Delivery Ratio or PDR, mea-sured as the ratio of the number of packets delivered totraffic sinks relative to the number of packets sent by thetraffic sources. However, in our scenarios, where nodesdie due to the depletion of their energy source, this metriccan be misleading. As all our nodes are energy-constrained,including traffic sources and sinks, a routing protocol thatwould assure that traffic sources deplete their energy firstwould thus achieve high PDR metrics, even though thenetwork itself only delivered very few packets in total.We therefore evaluate the various protocol variants by

14500

15000

15500

16000

16500

17000

17500

18000

18500

19000

19500

0.2 0.14 0.09 0.04

Packet Interarrival Time (seconds)

# Pa

cket

s

Modified Routing (Ideal)

Modified Routing (Real)

Modified Routing/MPR Selection (Ideal)

Modified Routing/MPR Selection (Real)

Fig. 2. Protocol variant performance for mobile network scenarios.

12000

14000

16000

18000

20000

22000

24000

26000

28000

30000

0.2 0.14 0.09 0.04

Packet Interarrival Time (seconds)

# Pa

cket

s

EOLSR, Static (Ideal)EOLSR, Static (Real)EOLSR, Mobility (Ideal)EOLSR, Mobility (Real)

Fig. 3. Protocol variant performance for EOLSR, static and mobilescenarios.

760 T. Kunz, R. Alhalimi / Ad Hoc Networks 8 (2010) 755–766

counting the total number of packets successfully deliv-ered, similar to some of the results reported in [8]. As weobserved from our simulation results, this metric is closelyrelated to other often reported metrics: the protocol vari-ant that delivers more packets also typically causes nodesto die later and the network to partition later. The onecommon metric that we found was not impacted by energyaccuracy was packet latency.

Figs. 1 and 2 summarize the results for the two mobilityscenarios, plotting traffic load against protocol perfor-mance for the various protocol variants introduced above.In both figures, the solid lines indicate the performance forthe Modified Routing variation of OLSR, and the dashedlines show the performance for the Modified MPR/Routingvariant. The square symbol indicates the performance ofthe ideal protocol version, whereas the diamond showsthe performance of the realistic protocol version. For thestatic scenarios, shown in Fig. 1, there is a clear differencebetween the ideal and realistic version, with the realisticversion delivering from close to 250 packets at low trafficrates (high packet interarrival times) to roughly 1000 pack-ets fewer than the ideal version for the Modified Routingvariant. For the mobile scenarios, the total number of pack-ets delivered drops, see Fig. 2. This is consistent with manyother MANET protocol performance studies that found thatmobility has a detrimental impact on overall performance.However, similar to the static scenarios, the ideal versionoutperforms the realistic version for all traffic loads.

In Fig. 3, we provide additional evidence that the effectof inaccurate energy knowledge is not limited to our ver-sions of an energy-efficient version of OLSR. We imple-mented EOLSR, as described in [8], and ran repeatedsimulations under the same conditions as above. Again,the protocol version that has omniscient knowledge of anode’s energy level delivers more packets than the realisticversion. We therefore conclude that collecting energy met-rics such as a node’s residual energy level to utilize in anenergy-efficient routing protocol, through regular protocolmessage exchanges, does not result in the best possibleprotocol performance that could be achieved. Dependingon the traffic rate and network mobility, an ideal protocolversion, having omniscient access to the current, instanta-neous energy values when making routing decisions, typi-cally achieves a much superior protocol performance,delivering up to 20% more packets. This performance gap

18000

19000

20000

21000

22000

23000

24000

25000

26000

27000

28000

0.2 0.14 0.09 0.04

Packet Interarrival Time (seconds)

# Pa

cket

s

Modified Routing (Ideal)

Modified Routing (Real)

Modified Routing/MPR Selection (Ideal)

Modified Routing/MPR Selection(Real)

Fig. 1. Protocol variant performance for static network scenarios.

motivates the remainder of the paper, where we explorethe accuracy of the collected information in more depth,as well as proposed strategies that allow nodes to have amore accurate view of energy levels at other nodes.

5. Analyzing QoS state information inaccuracy

To quantify the accuracy of state information in differ-ent conditions, we ran extensive simulations using theNS2 simulator with the OOLSR implementation of OLSRprovided by the Hipercom project (NS2 version 2.27 withOOLSR version 0.99.15) [9]. All simulations use the param-eters specified in Table 1, unless noted otherwise.

For each sample point, 10 random network snapshotsare generated. The simulation results presented are anaverage over these 10 scenarios. The same set of 10 scenar-ios are used for all simulations for a given sample point,hence the different parameters are evaluated under identi-cal conditions.

In the following, we describe the simulation conditionsthat we generally used in our simulation experiments. Ineach run, we have 20 unicast traffic flows between ran-domly selected node pairs. Each source sends 128 bytesCBR packets at different intervals. The simulations aredone with four different intervals to study the effect oflow, medium and high traffic rates on the QoS metric.

Table 1Common simulation parameters.

Propagation model TwoRayGround

Simulator parametersNetwork type IEEE 802.11Transmission range 250 mMobility model Static networkMAC bandwidth 11 Mpbs

Scenario parametersTopology area 1000 � 1000Number of nodes 50Simulation time (s) 200 s

Energy model specificationsInitial energy (J) 1000Transmission power (Watt) 1.4Receiving power (W) 1.0Idle power (W) 0.83

0

1

2

3

4

5

6

7

8

0.2 0.14 0.09 0.04

Packet Interarrival Time (seconds)

Inac

cura

cy (J

oule

s)

Fig. 4. Energy overall inaccuracy level under different packet interarrivaltimes with default OLSR parameters.

T. Kunz, R. Alhalimi / Ad Hoc Networks 8 (2010) 755–766 761

The intervals are 0.2, 0.14, 0.09, and 0.04 s – an interval of0.2 means a node will send a packet every 0.2 of a second(i.e., 5 packets per second). At the lower intervals, in partic-ular for 0.04 s, a significant number of packets (data andcontrol packets) will be lost due to network congestion,testing the performance under stressful overload scenarios.These may not be representative of long-term sustainedtraffic, but often occur in MANETs due to the link band-width limitations and the bursty nature of data traffic.The energy model values are based on the study conductedin [21].

During the simulation, a snapshot for the whole topol-ogy is taken every second. It contains information such as:

� Which other nodes a node can hear and what it believesto be their energy levels. Node n2 is said to be heard bynode n1 if there exists a route from n1 to n2, i.e., n1 hasa routing table entry for n2.

� If a node n2 is heard by node n1, how many hops away isit from node n1.

� The actual energy level for each node.

Also during the simulation, we keep records of all Helloand TC messages sent and received. We are interested indetermining the average overall inaccuracy level. We de-fine this as the average difference between a node’s actualenergy level and what other nodes believe its energy levelis. More formally, overall inaccuracy level is calculated as:for each pair of nodes (n1, n2) in the network such that n1can hear n2, the sum of the absolute difference betweenthe actual energy level of n2 and what n1 believes to bethe energy level of n2, for all time points (every secondof the simulation) divided by the total number of pairs.This is done for 10 scenarios and the overall inaccuracy le-vel is the average over the 10 scenarios. We ignore the first50 s of the simulation to have the network stabilized and toignore transient startup conditions. We also are only con-sidering pairs of nodes (n1, n2) such that n1 can hear n2.During the simulation and due to message loss and delays,some nodes get temporarily disconnected from othernodes. Therefore they are not considered part of the net-work as they are not used for routing. For overall inaccu-

racy level calculation, only visible nodes (i.e., nodes towhom a valid routing table entry exists) are considered.

In OLSR and its energy-efficient variants, only MPRs areselected as intermediate nodes on a path. Therefore, it mayappear preferable to only determine average inaccuracylevels for MPRs. However, all nodes in a network are poten-tially MPRs. In addition, in energy-efficient variants ofOLSR such as EOLSR, the perceived residual energy levelof one- and two-hop neighbors impacts the MPR selection.Inaccurate information may lead to poor MPR choices, evenbefore routes are being constructed over these MPRs. Wetherefore choose to measure the inaccuracy level for allnodes.

Fig. 4 shows the overall inaccuracy level under the dif-ferent traffic rates using the default OLSR parameters (i.e.,Hello-interval 2, TC-interval 5, MPR-coverage 1 and TC-redundancy 0). As expected, traffic does introduce a con-siderable level of inaccuracy to the network. And as thetraffic rate increases, the level of overall inaccuracy in-creases. As the traffic rate increases, the network becomesmore congested. Since network buffers have limited capac-ity in terms of storage and processing of arriving packets,this significantly affects the performance of the network,causing long delays and packet loss as packets will be wait-ing in the queues for processing or will be dropped due tooverflow. And as nodes send/receive traffic at a higher rate,they consume energy at a higher rate. Consequently, infor-mation available to the nodes in the network becomes out-dated and no longer accurate.

These observations are confirmed by our experiments.Table 2 shows the average nodal energy consumption, Hel-lo/TC message delay and Hello/TC message loss as well asaverage knowledge age under the different traffic rates.Hello/TC message loss is calculated as the percentage ofHello/TC messages received compared to the number of re-ceived messages using a packet interarrival time of 0.2 asbase case (as we are interested in what contributes to anincrease in knowledge inaccuracy, we look at changes rel-ative to this case). Average knowledge age is calculated as:for all pairs of nodes (n1, n2) in the network, the average ofhow old is the information n1 has about n2. At the highertraffic rates the TC message loss exceeds the Hello Messageloss: if a specific TC message does not reach some MPRs,they in turn will not rebroadcast the TC message, whichis also counted as a loss and included in these numbers.

Table 2Hello/TC message delay and loss, average knowledge age and nodal energy consumption.

Packetinterarrival time

Nodal energyconsumption (J/s)

Hello messagedelay (s)

TC messagedelay (s)

Hello messageloss (%)

TC messageloss (%)

Average knowledgeage (s)

0.2 0.926 0.016 0.114 – – 2.940.14 0.948 0.12 0.542 4 3 3.680.09 0.956 0.478 1.586 9 17 5.460.04 0.965 0.866 2.127 15 36 7.03

762 T. Kunz, R. Alhalimi / Ad Hoc Networks 8 (2010) 755–766

In contrast, Hello messages are propagated simply in the 1-hop neighborhood.

OLSR is very focused on overhead reduction. ThereforeOLSR default parameters are set to achieve an acceptableperformance (for a best-effort routing protocol) whilekeeping the overhead as low as possible. We investigatedif it is possible to tradeoff cost (overhead) to gain betterperformance (more accurate state information). We ana-lyzed the impact of sending more frequent Hello and TCmessages (by reducing Hello and TC intervals) as well asmore redundant topology information (by increasing TC-redundancy and MPR-coverage parameters). Choosing verysmall values for Hello and TC intervals will significantly in-crease the protocol overheads, in particular for TC mes-sages that are flooded throughout the network. Whilethis may be beneficial to the accuracy of the collected stateinformation, the increased control message overheadwould be quite detrimental to the data traffic and consumea non-trivial amount of energy. We therefore explored nei-ther very small Hello message intervals (less than 1 s) norsmall TC message intervals (less than 3 s).

As shown in Table 3, a 95% confidence interval is calcu-lated for each parameter under the different traffic rates.Increasing the number of protocol messages, includingboth Hello and TC messages, or increasing the amount ofinformation advertised improves the overall inaccuracy le-vel under low traffic rate. Under medium to high trafficrates, a 95% confidence interval test shows that the differ-ence between the overall accuracy levels under the differ-ent OLSR parameters is not statistically significant. Undermedium traffic rates (traffic intervals 0.14 and 0.09) we ob-serve a trend towards better inaccuracy levels when vary-ing the OLSR parameters. On the other hand, under thehighest traffic rate, the trend is towards less accurate en-ergy levels when varying the OLSR parameters. These re-sults are a direct consequence of the increased level ofcongestion in the network which results in high messageloss and delay and hence less accurate state information.

Table 3 shows that among the different OLSR parame-ters, under low traffic rates, improvement in inaccuracy le-vel is best achieved by a TC-interval of 3 with someimprovement achieved by lowering the Hello-intervalfrom 2 to 1. According to the OLSR specifications, the ratiobetween Hello-interval and TC-interval is 2 to 5. Thereforeusing a combination of TC-interval of 3 and Hello-intervalof 1 seems appropriate. Fig. 5 compares the overall inaccu-racy level using the default parameters for OLSR, called de-fault OLSR, versus using a combination of Hello-interval 1and TC-interval 3, keeping the other parameters un-changed (MPR-coverage 1 and TC-redundancy 0), calledHello1TC3 OLSR. Hello1TC3 OLSR improves overall inaccu-

racy level under low traffic rates. At high traffic rates, De-fault OLSR starts to outperform Hello1TC3 OLSR due tothe overhead added to the network (more frequent HELLOand TC messages).

The obvious approach of increasing the frequency ofprotocol messages or the amount of information adver-tised improves the energy overall inaccuracy level underlow traffic rates only. Therefore, we further analyzed theresults based on knowledge age to obtain insights of whatcan be done to improve inaccuracies under higher trafficrates or even improve it further under low traffic rates.As a first step, we correlated knowledge and inaccuracy le-vel. Knowledge age inaccuracy level represents the averagedifference between a node’s actual energy level and whatother nodes believe its energy level is, categorized byhow old the knowledge is. The knowledge between allpairs of connected nodes is categorized into 22 differentgroups based on how old the data is. The 1st group hasall pairs with knowledge that is 0 (inclusive) to 1 s old(exclusive). And the 2nd group has all pairs with knowl-edge that is 1 (inclusive) to 2 s old (exclusive) and so on.All nodes that have data which is above 21 s old are putin the last group since we observed that it is very rare tohave data that is older than 21 s. According to the OLSRimplementation with default parameter values, it is typicalto have data that is up to 6 s old which is the TC-intervaldefault (every 5 s) plus 1 s to propagate the message. Eachtuple in a node’s topology database will be expired after15 s (or three times the TC-interval), so a reasonable upperbound on knowledge age under not heavily congested net-work conditions is therefore 21 s.

Fig. 6 illustrates the results of knowledge age inaccu-racy level under the four different traffic rates and usingthe default OLSR parameters. It can be clearly seen fromthis figure that the older the knowledge about other nodesis, the less accurate their knowledge about the node’s en-ergy levels are. This trend is easily explained by the factthat the energy level is a monotonically decreasing metric.

Note that Fig. 6 only correlates age with inaccuracy; itdoes not detail how old, on average, the information at agiven node is. Table 2 above provides the average knowl-edge age under different packet interarrival times/trafficrates, as well as the average latency and loss rates for Helloand TC messages, using the default OLSR parameters. Theaverage knowledge age increases with the traffic rate, asdoes the loss of Hello and TC messages. In the simulationsetup, the OLSR control packets are given priority overthe data packets within a node, using a priority queue.Consequently, for low traffic rates, these messages havevery low latencies. Also, all broadcast packets are ran-domly jittered as described in the OLSR RFC [1] to avoid

T. Kunz, R. Alhalimi / Ad Hoc Networks 8 (2010) 755–766 763

collisions. As the traffic load increases, the control packetshave to contend for access to the channel with packets atthe head of the queue at neighboring nodes, increasingthe latency and loss rates due to collisions. We could ex-plore the use of mechanisms such as providing controlpackets priority access to the channel using IEEE 802.11e,reducing the packet latency and loss rates under highertraffic loads (albeit at a cost to the data packet delivery).However, even under low traffic rates, the average knowl-edge age is a few seconds, as information gets refreshedonly by periodic control messages. As Figs. 1–3 show, evenunder the lowest traffic load (highest packet interarrivaltime), the resulting knowledge inaccuracy has detrimentaleffects on the protocol performance.

Our results so far show that nodes have inaccurateinformation about the actual residual energy levels whenmaking routing decisions. Modifying the OLSR protocolparameters (such as increasing the Hello or TC messagerates) has very limited impact on this inaccuracy. Basedon these observations, the next section explores additionalways to increase energy level accuracy above and beyondmodifying the protocol parameters.

6. Improving accuracy

Based on the energy level knowledge age inaccuracy le-vel analysis in the previous section, the results show thatthe main source of inaccuracies is the existence of old data.A first approach to address this problem would be for anode to selectively probe nodes for which it holds oldinformation (above a certain threshold, say), hopefullyobtaining more recent (and therefore more accurate) stateinformation. In work reported elsewhere [22], we experi-mented with such an approach and found that it, in gen-eral, is not capable of achieving significant reductions ininaccuracy.

Alternatively, as energy level is a monotonicallydecreasing metric, a node can adjust ‘‘old” informationand predict the current value whenever it needs to deter-mine the residual energy level for a visible node (whenselecting MPRs or determining a routing path, for exam-ple). Our idea is therefore to have every node locally adjustnodes’ old energy levels based on their past energy con-sumption rate. We propose a Prediction mechanism inwhich each node locally extrapolates the expected energy

Table 3Overall inaccuracy level (in Joules) under different packet interarrival times with

Traffic interval OLSR parameters 0.2 0

Hello-interval 2 2.6225 [2.37,2.87] 31 2.213 [1.99,2.43] 2

TC-interval 5 2.6225 [2.37,2.87] 34 2.196 [1.95,2.43] 23 1.981 [1.76,2.2] 2

MPR-coverage 1 2.6225[2.37,2.87] 32 2.241 [2.05,2.43] 2

TC-redundancy 0 2.6225 [2.37,2.87] 31 2.283 [2.06,2.5] 32 2.143 [1.98,2.3] 3

level based on old (reported) energy levels and the energyconsumption rate for that node based on the most recenttwo reported values. For example, at second 51 of the sim-ulation, node 0 has an energy level of 958.581 J associatedwith node 1 and this knowledge is timestamped 47.5916 s.At second 52, the perceived energy level for node 1 (fromthe perspective of node 0) is 954.998 J, timestamped49.5884s. Node 0 computes the consumption rate of node1 as:

ð958:581 J�954:998 JÞ=ð49:5884 s�47:5916 sÞ¼1:7943 J=s:

To determine the perceived residual energy level for node1 at second 52, node 0 adjusts the last reported energy le-vel of node 1 by the consumption rate and elapsed time:

Predicated perceived residual energy of node 1:954.998 J � (1.7943 J/s � (52 s � 49.5884 s)) = 950.6707 J.

If no prediction is possible, as no consumption rate isknown yet, the last reported energy level will be usedwithout adjustment.

A drawback of the Prediction algorithm is the need towait for two different perceived value readings, so a con-sumption rate can be calculated and used to adjust the per-ceived values. Table 4 shows how often an adjustmenttakes place under different traffic rates. These results cap-ture how often we are successful in adjusting the last re-ported residual energy level as a percentage of thenumber of times we need knowledge about a node’s resid-ual energy level (to select MPRs, determine routes, or re-port perceived energy levels for the purpose of collectingstatistic). Under high traffic loads, adjustments happen lessrarely. Protocol control messages are lost/delayed, and as aresult nodes will not ‘‘hear” other nodes. After a node isdeemed unreachable, we go through the startup phaseagain, where we need at least two successive reports tobe able to calculate a consumption rate. We therefore pro-pose the Smart Prediction algorithm which is an enhancedversion of the prediction algorithm so that adjustmentstake place almost all the time.

In the Smart Prediction algorithm, for every pair ofnodes (n1, n2), if n2’s consumption rate is not yet known,n1 adjusts the perceived value of n2’s residual energy levelbased on the average of all known consumption rates forother nodes. If n1 knows not a single consumption ratefor other nodes, it adjusts n2’s perceived energy levelbased on its (n1’s) consumption rate. Using all known

different OLSR parameters.

.14 0.09 0.04

.4287 [2.96,3.9] 5.1793 [4.8,5.56] 6.7602 [5.94,7.58]

.956 [2.47,3.44] 4.846 [4.36,5.33] 6.992 [6.16,7.82]

.4287 [2.96,3.9] 5.1793 [4.8,5.56] 6.7602 [5.94,7.58]

.978 [2.49,3.46] 5.014 [4.59,5.44] 6.774 [5.9,7.64]

.865 [2.38,3.35] 5.097 [4.66,5.53] 7.004 [6.2,7.8]

.4287 [2.96,3.9] 5.1793 [4.8,5.56] 6.7602 [5.94,7.58]

.888 [2.47,3.31] 4.718 [4.22,5.21] 6.925 [5.99,7.86]

.4287 [2.96,3.9] 5.1793 [4.8,5.56] 6.7602 [5.94,7.58]

.062 [2.6,3.52] 5.014 [4.51,5.52] 6.81 [5.98,7.64]

.005 [2.51,3.5] 4.89 [4.47,5.31] 6.79 [6.01,7.57]

0

1

2

3

4

5

6

7

8

0.2 0.14 0.09 0.04

Packet Interarrival Time (seconds)

Inac

cura

cy (

Joul

es)

Hello1TC3 OLSR Default OLSR

Fig. 5. Overall inaccuracy level of default OLSR vs. Hello1TC3 OLSR.0

1

2

3

4

5

6

7

8

0.2 0.14 0.09 0.04

Packet Interarrival Time (seconds)

Inac

cura

cy (J

oule

s)

Default OLSR Prediction Smart Prediction

Fig. 7. Overall inaccuracy level using default OLSR vs. prediction vs. smartprediction.

764 T. Kunz, R. Alhalimi / Ad Hoc Networks 8 (2010) 755–766

nodes’ consumption rates eliminates the domination ofoutliers and ensures closeness to the actual consumptionrate, assuming that nodes are somewhat homogeneous inthe energy characteristics of their wireless cards.

According to Fig. 7, the Prediction algorithms improvethe overall inaccuracy level under different traffic rates.The improvement under higher traffic rates is not as highas it is under lower traffic rates. For an adjustment to takeplace, a node must have received two different reportedvalues. But under high traffic rates, due to message lossand delays, the percentage of times adjustments take placedecreases, as shown in Table 4. Since the Smart Predictionalgorithm addresses the problem of not being able to ad-just the perceived energy level value all the time, itachieves much better performance in terms of overall inac-curacy level, especially under higher traffic rates. Both thePrediction and the Smart Prediction algorithms outperformthe Default OLSR protocol. At the same time, the Smart Pre-diction algorithm outperforms the Prediction algorithm inimproving the overall inaccuracy level.

7. Mobile scenarios and heterogenous interfaces

The previous results were all based on static scenariosand under the assumption that wireless nodes have thesame radio (or at least radios with identical energy con-sumption characteristics). We also performed additionalexperiments where we allowed nodes to move and where

0

5

10

15

20

25

30

[0.0,1

.0[

[1.0,2

.0[

[2.0,3

.0[

[3.0,4

.0[

[4.0,5

.0[

[5.0,6

.0[

[6.0,7

.0[

[7.0,8

.0[

[8.0,9

.0[

[9.0,1

0.0[

[10.0,

11.0[

[11.0,12

.0[

[12.0,

13.0[

[13.0,

14.0[

[14.0,15

.0[

[15.0,

16.0[

[16.0,17

.0[

[17.0,

18.0[

[18.0,

19.0[

[19.0,

20.0[

[20.0,21

.0[

21 an

d abo

ve

Knowledge Age

Inac

cura

cy (

Joul

es)

Packet Interarrival Time 0.2 Packet Interarrival Time 0.14Packet Interarrival Time 0.09 Packet Interarrival Time 0.04

Fig. 6. Knowledge age inaccuracy level under different packet interarrivaltimes and default OLSR parameters.

the energy characteristics of different nodes/radios weremore heterogeneous. For mobility, we used the mobilityscenario described in Section 4. The accuracy of the resid-ual energy level metric, using only protocol parameters(Default OSLR), Prediction, and Smart Prediction, areshown in Fig. 8.

These results are collected assuming node homogene-ity, i.e., each mobile node uses a radio with identical energyconsumption for sending/receiving, and idle time. The fig-ure, similar to Fig. 7, show that the smart prediction strat-egy produces by far the most accurate results. The resultsare nearly identical for static and mobile scenarios, indicat-ing that mobility has little impact on the performance ofthe prediction methods.

To model more heterogeneous nodes, we assigned ran-dom values to the three energy model parameters asfollows:

– Transmission power: uniformly distributed between 1.2and 1.6 W.

– Receive power: uniformly distributed between 0.8 and1.1 W.

– Idle power consumption: uniformly distributed between0.3 and 0.7 W.

As a result, we model radios where sending is morecostly (in energy terms) than receiving, and idle powerconsumption is lower but still non-negligible. As each nodegets assigned a different set of parameters in the simula-tion setup, no two nodes will (with high probability) havethe same radio. We also varied the initial nodal energy lev-els, randomly assigning a value uniformly distributed be-tween 500 and 1000 J.

Table 4Percentage of times adjustments take place under different packet inter-arrival times.

Packet interarrival time Number of adjustments (in %)

0.2 970.14 950.09 910.04 82

Default OLSR Prediction Smart Prediction

0

1

2

3

4

5

6

7

8

0.2 0.14 0.09 0.04

Packet Interarrival Time (seconds)

Inac

cura

cy (J

oule

s)

Fig. 8. Overall inaccuracy level, mobile scenarios, homogeneous radios.

0

1

2

3

4

5

6

7

8

9

10Default OLSR Prediction Smart Prediction

0.2 0.14 0.09 0.04

Packet Interarrival Time (seconds)

Inac

cura

cy (J

oule

s)

Fig. 9. Overall inaccuracy level, static scenarios, heterogeneous radios.

T. Kunz, R. Alhalimi / Ad Hoc Networks 8 (2010) 755–766 765

Figs. 9 and 10 show the resulting inaccuracies for thethree different options of determining residual energy lev-els. Similar to the homogeneous case, prediction outper-forms Default OLSR and smart prediction againoutperforms prediction. These results are true for bothmobility scenarios. Compared to Figs. 7 and 8, overall inac-curacy is slightly higher, due to the higher variability in theenergy characteristics of the radios.

0

1

2

3

4

5

6

7

8

9

10Default OLSR Prediction Smart Prediction

0.2 0.14 0.09 0.04

Packet Interarrival Time (seconds)

Inac

cura

cy (J

oule

s)

Fig. 10. Overall inaccuracy level, mobile scenarios, heterogeneous radios.

8. Conclusions and future work

In this work, we used the Optimized Link State Routing(OLSR) protocol as the underlying wireless network rout-ing protocol. OLSR can be modified to become a more en-ergy-aware routing protocol. However, the potentialimprovements are limited by the accuracy of the collectedenergy-related metric (here, a node’s residual energylevel).

We quantified the state information accuracy (energylevel) under different traffic rates. The results showed thatas traffic rate increases, the overall inaccuracy level in-creases. Tuning the OLSR protocol parameters did not havea noticeable impact on overall inaccuracy level.

We proposed two techniques to reduce energy levelinaccuracies, Prediction and Smart Prediction. Under thePrediction technique, a node’s energy level is adjustedbased on its past behavior (its own consumption rate).Smart Prediction is a modified version of the Predictiontechnique such that, if no consumption rate can be deter-mined for a node, its energy level is adjusted based onthe average of all known consumption rates for othernodes. The results show that both Prediction and SmartPrediction outperform the Default OLSR (OLSR with QoS-related state propagated, and using default parameters).Moreover, Smart Prediction outperforms Prediction sinceenergy level adjustments take place all the time. In addi-tion, the overheads associated with Prediction and SmartPrediction are exactly the same as in Default OLSR sinceno extra messages or fields are required. Both predictiontechniques work well under different mobility levels andfor heterogeneous wireless radios.

The proposed techniques work well for energy level,which is a monotonically decreasing metric. Other rout-ing-relevant metrics such as queue length (for load-bal-anced routing), available link bandwidth (for QoSrouting), etc., are not monotonic. In [22], we explored theinaccuracy of a node’s outbound queue length, and showedthat, similar to residual energy level, increased traffic loadswill result in increasingly inaccurate information, and thata simple technique such as explicitly probing for old infor-mation does not improve the information accuracy. Wehave not yet verified that this inaccurate information willindeed lead to poor protocol performance, but expect thisto be the case. Once this is done, we then need to identifyadditional techniques to increase the collected informationaccuracy.

Acknowledgment

This work was supported by the National Sciences andEngineering Research Council of Canada.

References

[1] T. Clausen, P. Jacquet, Optimized link state routing protocol, IETF RFC3626, 2003.

[2] J.H. Chang, L. Tassiulas, Energy conserving routing in wireless ad-hocnetworks, in: Proceedings of IEEE INFOCOM, Tel Aviv, Israel, vol. 1,March 2000, pp. 22–31.

766 T. Kunz, R. Alhalimi / Ad Hoc Networks 8 (2010) 755–766

[3] S. Singh, M. Woo, S. Raghavendra, Power-aware routing in mobile adhoc networks, in: Proceedings of the ACM Mobile Computing andNetworking Conference, Dallas, TX, October 1998, pp. 181–190.

[4] C.-K. Toh, Maximum battery life routing to support ubiquitousmobile computing in wireless ad hoc networks, IEEECommunications Magazine (2001) 2–11.

[5] T. Clausen, P. Jacquet, L. Viennot, Investigating the impact of partialtopology in proactive MANET routing protocols, in: Proceedings ofthe 5th International Symposium on Wireless Personal MultimediaCommunications, vol. 3, October 2002, pp. 1374–1378.

[6] T. Plesse et al., OLSR performance measurement in a military mobilead-hoc network, in: Proceedings of the 24th InternationalConference on Distributed Computing Systems Workshops, October2004, pp. 704–709.

[7] T. Kunz, Energy-efficient variations of OLSR, in: Proceedings of theInternational Wireless Communications and Mobile ComputingConference (IWCMC 2008), Crete, Greece, August 2008, pp. 517–522.

[8] S. Mahfoudh, P. Minet, EOLSR: an energy efficient routing protocol inwireless ad hoc sensor networks, Journal of InterconnectionNetworks 9 (4) (2008) 389–408.

[9] Hipercom Project, OOLSR Implementation of the OLSR OptimizedLink State Routing Protocol, November 2004. <http://hipercom.inria.fr/oolsr/>.

[10] A. Tonnesen, Implementing and Extending the Optimized Link StateRouting Protocol, Master Thesis, Department of Informatics,University of Oslo, August 2004.

[11] Y. Ge, T. Kunz, L. Lamont, Proactive QoS routing in ad-hoc networks,in: Proceedings of the 2nd International Conference on Ad-HocNetworks and Wireless, Montreal, Canada, October 2003.

[12] P.E. Villanueva-Pena, T. Kunz, P. Dhakal, Extending networkknowledge: making OLSR a quality of service conductiveprotocol, in: Proceedings of the International Conference onCommunications and Mobile Computing, Vancouver, Canada, July2006, pp. 103–108.

[13] N. Ghanem et al., New energy saving mechanisms for mobile ad-hocnetworks using OLSR, in: Proceedings of the 2nd ACM InternationalWorkshop on Performance Evaluation of Wireless Ad Hoc, Sensor,and Ubiquitous Networks, October 2005, pp. 273–274.

[14] Z. Guo, B. Malakooti, Energy aware proactive MANET routing withprediction on energy consumption, in: Proceedings of the 2007International Conference on Wireless Algorithms, Systems andApplications, August 2007, pp. 287–293.

[15] A. Benslimane et al., Energy power-aware routing in OLSR proto-col, in: Proceedings of the First Mobile Computing and WirelessCommunication International Conference (MCWC 2006), September2006, pp. 14–19.

[16] Q. Han, N. Venkatasubramanian, Information collection services forQoS – aware mobile applications, IEEE Transactions on MobileComputing 5 (5) (2006) 518–535.

[17] G. Apostolopoulos, R. Guerin, S. Kamat, S. Tripathi, Quality of servicebased routing: a performance perspective, in: Proceedings of ACMSIGCOMM, Vancouver, Canada, vol. 28 (4), October 1998, pp. 17–28.

[18] R.A. Guerin, A. Orda, QoS routing in networks with inaccurateinformation: theory and algorithms, IEEE/ACM Transactions onNetworking 7 (3) (1999) 350–364.

[19] A. Shaikh, J. Rexford, K.G. Shin, Evaluating the impact of stale linkstate on quality-of-service routing, IEEE/ACM Transactions onNetworking 9 (2) (2001) 162–176.

[20] S. Corson, J. Macker, Mobile Ad hoc Networking (MANET): RoutingProtocol Performance Issues and Evaluation Considerations, RFC2501, January 1999. <http://www.ietf.org/rfc/rfc2501.txt>.

[21] L.M. Feeney, M. Nilsson, Investigating the Energy consumption of awireless network interface in an ad hoc networking environment, in:Proceedings of IEEE INFOCOM, Anchorage, AK, vol. 3, April 2001, pp.1548–1557.

[22] T. Kunz, R. Alhalimi, Load-balanced routing in wireless networks:state information accuracy using OLSR, in: Proceedings of the 3rdIEEE International Conference on Wireless and Mobile Computing,Networking and Communications (WiMob 2007), New York, USA,October 2007.

Thomas Kunz is currently a Professor inSystems and Computer Engineering at Carle-ton University, Canada. He heads the MobileComputing group, researching wireless net-work architectures (MANETs, wireless meshnetworks, wireless sensor networks), networkprotocols (routing, Mobile IP, QoS support),and middleware layers for innovative wirelessapplications. Professor Kunz authored or co-authored over 130 technical papers, receiveda number of awards, and has served on over40 TPCs of international conferences and

workshops in the mobile and wireless domain.

Rana Alhalini completed a Masters Degree inComputer Science at Carleton University andis currently working at Rockwell Collins Sys-tems. Her research interests are in Mobile Ad-Hoc Networking, more specifically routingprotocols and extensions to support QoSrouting.