1

An Interference-Aware Admission ControlDesign for Wireless Mesh Networks

Devu Manikantan Shila and Tricha AnjaliDepartment of Electrical and Computer Engineering

Email: {dmanikan, anjali}@iit.edu

Abstract—With the increasing popularity of wireless mesh networks(WMNs), the demand for multimedia services encompassing VoIP, multi-media streaming and interactive gaming is increasing rapidly. The real-time services require stringent Quality of Service (QoS) guarantees foreffective communication, as compared tobest-effort services (for e.g, filetransfer) that are tolerant to changes in bandwidth and delay. While a lotof research has dealt with providing QoS support forreal-time servicesin traditional wired networks, the shared and broadcast nature of thewireless medium necessitates the design of new solutions for multi-hopwireless networks. In multi-hop wireless networks, unlikewired networks,the communication from one node will consume the bandwidth of theneighboring nodes and hence the shared bandwidth can be easily over-utilized. Therefore, to provide an acceptable level of QoS for the real-time services, it is necessary to control the utilization of the sharedbandwidth. In this paper, we propose an efficient admission controlalgorithm, Interference-aware Admission Control (IAC), for multi-hopmesh networks which aims at preserving the QoS for all the admittedflows by employing a low overhead threshold mechanism. We describeseveral alternatives for the design of IAC and compare the performanceof these alternatives using simulation results. The simulation results alsodemonstrate that IAC ensures high throughput, low packet loss and delayfor all admitted flows.

Index Terms—wireless mesh networks, QoS, admission control, intra-flow interference, inter-flow interference

I. I NTRODUCTION

Recent advances in wireless network communications have lead tothe development of Wireless Mesh Networks (WMNs)[1], [2], [3],a promising technology that has the potential to provide wirelessservices in locations where there is currently little or no infrastructure.In WMNs, a collection of wireless mesh routers, usually employingIEEE 802.11 based commodity hardware operating in ad hoc mode,provides network access to the wireless clients. However, commu-nications between mesh routers are realized by radio transmissionsusually in a multi-hop fashion. One or more mesh routers in the net-work are connected to the Internet and serve as gateways to provideInternet connectivity for the entire mesh network. The existence of awireless backhaul provides many advantages for mesh networks likemarket coverage, scalability and low upfront cost.

As WMNs grow in popularity with numerous public and privatedeployments [4], [5], [6], the demand for multimedia services, suchas Voice over IP (VoIP), Video on Demand (VoD) and interactivegaming, along with traditional data services is also increasing rapidly.These real-time services have stringent Quality of Service (QoS)requirements (for e.g, low packet loss ratio and delay) for effectivecommunication as compared tobest-effortservices that are tolerantto changes in bandwidth and delay [7], [8], [9], [10], [11]. Inspite ofthe multiple advantages, it is fairly difficult to maintain a desiredlevel of QoS for multimedia applications in a multihop networklike WMN. This is attributed to many factors [1], [2], [3]: (a)Unlike wired networks, the devices in mesh networks communicatevia radio transmissions and hence, packet losses can occur due toerror prone nature of wireless channel; (b) The interference amongconcurrent wireless transmissions either on a multihop path (intra-flow interference) or on multiple paths in proximity (inter-flow

interference) have an adverse impact on the performance of real-time flows. We have demonstrated the results of this impact usingsimulations in section II. Therefore, while designing algorithms forwireless mesh networks, one should take into account both the intra-flow and the inter-flow interference among wireless links.

The currentmodus operandiof QoS provisioning in wired net-works is not directly applicable to WMNs because of the differingnature of wired and wireless networks. Since WMNs share commoncharacteristics with the ad hoc networks, the QoS solutions designedfor ad hoc networks can be applied to WMNs. Nevertheless, the staticnature of mesh routers and the mobile nature of client nodes impliesthat the solutions for ad hoc networks may not be suitable for WMNs.More specifically, new solutions incorporating special characteristicsof WMNs need to be developed.

Generally, wireless networks have limited bandwidth resources anddue to the shared nature of the wireless medium, communicationfrom one node may use the bandwidth resources of the neighboringnodes which in turn leads to over-utilization of resources [12], [13],[14]. Therefore, it is critical to employ admission control for theprovisioning of QoS guaranteed services in wireless networks becauseit can avoid the system resources from being overused. In this paper,we propose an admission control algorithm known asInterference-aware Admission Control(IAC) for single channel wireless meshnetworks. IAC takes into account both the intra-flow and inter-flow interference effect to make admission decisions. Therefore, anew flow is admitted only if there is enough local and neighboringbandwidth resources to support the flow. Thereby, IAC ensures thatthe shared resources are not overused and preserves the QoS of alladmitted flows. The main contributions of this paper are given asfollows:

• We propose adynamic, distributedadmission control algorithmwhich adapts to the traffic conditions and takes into considera-tion the available bandwidth information of interfering neighborsin making admission decisions.

• We present a threshold based scheme known asdual-thresholdto mitigate the overhead caused by the bandwidth informationexchanges while guaranteeing a desired level of QoS for alladmitted flows in the network. The scheme also ensures that thechannel resources are efficiently utilized. We call this optimizedversion of IAC algorithm as IAC/O.

• We propose two novel approaches for available bandwidthallocation and ensure that the shared wireless bandwidth is notover-utilized and the quality of all existing flows are preserved.

The remainder of the paper is organized as follows: In section II, wediscuss the motivations behind our work by stressing the necessityof a minimum overhead admission control in multihop networks.In section III, we discuss the related work for admission controlin wired and wireless networks. Section IV addresses the design ofthe proposed admission control algorithm in detail. In section V,we discuss the implementation of the proposed algorithm in thecontext of AODV routing protocol [15]. In section VI, we present

2

0

1

2

4

5

6

8

9

10

12

13

14

250m

500m

3 7 11 15

f1 f2 f 3

Fig. 1. 4x4 Grid Simulation Topology. Three flows, 1:(2-3), 2:(6-7), 3:(14-15), are established.

an analysis of thedeterministicandadaptivedual-threshold scheme.Simulation results and their analysis are discussed in section VII. Insection VIII, we present the discussions and future enhancements.Finally, section IX concludes this paper.

II. EXAMPLE AND MOTIVATION

In mesh networks, due to the shared nature of the wireless medium,nodes compete for channel bandwidth not only with the neighborsin their communication range but also with the nodes in theircarrier sensing range. As a result, the shared channel bandwidthcan be easily over-utilized which in turn affects the quality of theexisting flows. To capture the available bandwidth of the wirelessnetwork accurately [12], we need to consider two types of bandwidth:local available bandwidthandneighbor available bandwidth. Localavailable bandwidth is the amount ofsparebandwidth observed bythe given node whereas neighbor available bandwidth is the amountof sparebandwidth that a given node can consume without affectingthe quality of the existing flows in its transmission and carrier sensingrange. Therefore, the amount of bandwidth available for a givennode is theminimumof the local available bandwidth and neighboravailable bandwidth. To further illustrate the importance of the localand neighbor available bandwidth in wireless networks, we show asimple simulation using Ns2(v2.29)[16] as follows.

A. Example

As shown in figure 1, we used a grid topology of 16 mesh nodes.The MAC layer protocol is IEEE 802.11 CSMA/CA with 250m radiotransmission range and 550m carrier sensing range. The bandwidthof the wireless channel is set to 2 Mb/s. Three CBR flows (1, 2,3) of 545Kb/s are established between nodes 2 and 3, nodes 6 and7, and nodes 14 and 15 respectively. Nodes 2 and 14 are the directand carrier sensing neighbors of node 6 whereas nodes 2 and 14are out of the contention range. Each flow requires about 46.5% ofthe channel bandwidth due to the MAC layer (RTS/CTS/DATA/ACKhandshake) overhead [13], [22]. At time t =1 second, node 2 initiatesflow 1 to node 3. At t =2 seconds, node 6 initiates flow 2 to node7. Finally, at time t =3 seconds, node 14 initiates flow 3 to node15. Figures 2 and 3 shows respectively the throughput and delayof each flow over time. The results of the performance evaluationcan be explained as follows: Nodes 2 and 6 are direct neighborsand hence after flow 1 starts, the local available bandwidth at thesenodes is reduced to 53.5% of channel bandwidth. Since node 6 isa direct and carrier sensing neighbor of nodes 2 and 14, initiationof flow 2 consumes the bandwidth of nodes 2, 6, and 14. Thisreduces the local available bandwidth for these nodes to 7%, 7%and 53.5% of channel bandwidth respectively. At time t =3 seconds,node 14 admits the flow 3 for any one of the following reasons:(a) an admission control protocol is not employed; (b) an admissioncontrol protocol is employed. However it only considers the localavailable bandwidth (53.5% of channel bandwidth) and not the

neighbor available bandwidth (7% of channel bandwidth). What willhappen if node 14 admits the flowf3? The local available bandwidthof node 6 (7%) is much less than the consumption of flow 3 (46.5%).Therefore, when node 14 starts flow 3, the contention from it actuallydecreases the throughput of flow 2 by 8% (see fig. 2) and increasesthe delay of flow 2 from 0.05s to 0.45s (see fig. 3). Since thereal-time flows have stringent QoS requirments (for e.g., the delay foran acceptable VoIP flow should be less than 200ms according to theETSI recommendation [7], [8], [9] and the loss rate should be nomore than 1%), it is easy to conclude that the increased delay (0.05sto 0.45s) and reduced throughput of flow 2 will have an adverseimpact on the perceived quality of the flow.

0

100

200

300

400

500

600

700

1 2 3 4 5 6 7 8 9Time (Seconds)

Th

rou

gh

pu

t (k

bp

s)

Flow 1

Flow 2

Flow 3

Fig. 2. Throughput of three flows (1, 2, 3)vs Time. RTS/CTS mechanismis enabled.

0

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

1 2 3 4 5 6 7 8 9

Time (seconds)

Dela

y

Flow 1

Flow 2

Flow 3

Fig. 3. Delay of three flows (1, 2, 3)vs Time. RTS/CTS mechanism isenabled.

B. Motivation

To preserve the quality of existing flows in a multihop networklike WMN, one should take into account two kinds of interference:intra-flow interferenceand inter-flow interference1. Though our workis closely related to [12], the major difference of the proposedalgorithm, IAC, stems from mitigating the overhead caused by thebandwidth information exchanges seen in CACP [12], MARIA [17]while guaranteeing a desired level of QoS for all admitted flows inthe network. The issue is so important because the exchanges ofinformation proposed in [12] or [17] may cause high overhead inthe network and in turn consume a part of the bandwidth availablefor data transmission. To achieve the objective of reducing overhead,we propose novel dual-threshold scheme and event-driven messageexchange to share the bandwidth information in the neighborhood,which indeed is the motivation behind this work.

III. R ELATED WORK

In this section, we discuss several existing admission controlalgorithms in the area of wired, ad hoc and wireless mesh networks.Traditional approaches like IntServ/RSVP and DiffServ [11] for

1Inter-flow interference is the interference caused when thenodes on theadjacent path contend with each other for the channel bandwidth (for e.g.,flows 2 and 3 in fig. 1) whereas intra-flow interference is the interferencecaused when the nodes on the path of the same flow contend with each otherfor the channel bandwidth.

3

QoS provisioning on the internet do not work well for wirelessnetworks due to the inherent distinctions between the wired andwireless networks. For example, these approaches fail to address thechallenges involved in estimating the available bandwidth required bythe flow in a shared wireless medium. INSIGNIA [23] and SWAN[25] are both protocols that provide QoS guarantees by controlling thetraffic in ad hoc network. INSIGNIA relies on in-band signaling bypiggybacking the control information on data. This in-band signalingallows INSIGNIA to quickly re-establish flow state when topologychanges occur. SWAN, an alternative to INSIGNIA, is characterizedby its improved scalability. Here, the per-flow information is notstored in the intermediate nodes and therefore the protocol avoidscomplex signaling which makes the system more simple and scalable.However, the downside of both these schemes is that neither takesinto account the fact that the communication from one node mayconsume the resources of the neighboring nodes in addition to itslocal resources.

TAC [27] and AAC-AODV [29] implement flow admission controlbased on intra-flow interference. However, both these schemes donot give enough attention to inter-flow interference and thus, failto estimate the neighbor available bandwidth. In [24], [28], theauthors rely on interference-capacity model to design an admissioncontrol algorithm for VOIP calls in mesh network. The algorithmperforms admission control by monitoring the number of calls anddetermines whether the maximum capacity is reached. When thecall limit reaches the maximum, the admission controller will rejectfurther call requests. However, the approach fails to consider theload in the network. Since the degree of interference caused by eachnode depends on the traffic carried by the flow, merely monitoringthe calls can fail to accurately model the interference and canlead to erroneous admission decisions. Additionally, the authors usethe interference-capacity model obtained from a chain topology toachieve call admission control in generic random topology. However,such approximation schemes can not accurately model the inter-flowand intra-flow interference in a mesh network and thus can leadto admission of calls beyond the network capacity. In [18], authorspropose an admission control algorithm for multi-constrained, multi-level QoS routing in wireless mesh networks. Their design employstwo phases, the QoS performance index estimation phase, and multi-level QoS and GoS resource management phase to efficiently managethe resources among existing and new flows.

In MARIA [17], the authors characterize interference in wirelessmesh networks using the well-known conflict graph based model. Inthis approach, the nodes exchange their flow information periodicallyusing high power HELLO messages to determine their remainingavailable bandwidth based on the maximum clique constraints. Themain drawback of MARIA is that these periodic exchanges ofinformation using extended power cause high overhead in the networkand in turn consume a part of the bandwidth available for datatransmission. CACP [12] addresses admission control for ad hocnetworks and factors both the effect of intra-flow and inter-flowinterference. Nonetheless, CACP has the following drawbacks: (a)CACP supports only source routing protocols like DSR [21]; (b)CACP determines the available bandwidth of all the nodes within thetransmission and carrier sensing range by sending a query messageto these nodes. If any node that receives the query does not haveenough resources to support the flow, it sends a rejection message.This query process is performed on a hop-by-hop basis during theroute discovery phase. However, this process is time consuming andimposes a high message overhead in the network which in turnconsumes a part of the bandwidth available for data packet delivery[14]. PAC [14] performs admission control by relying on a passiveapproach to determine the available bandwidth at the interfering

neighbors. Although, this approach has lower overhead, the processof estimating available bandwidth is conservative. Additionally, PACdoes not give enough attention to intra-flow interference. As opposedto these approaches, in [19], authors present a statistical connectionadmission control framework to estimate the network resource fora pair of ingress-egress nodes and make admission decision basedon this estimated result. Their approach ensures a low signallingoverhead based admission control algorithm for wireless networks.

Some other notable works in the area of admission control algo-rithm in multihop networks are [34]-[37]. In the area of admissioncontrol algorithms for multi-channel wireless networks [34], [36],for instance authors in [34] propose a QoS routing for multi-radiomulti-channel IEEE 802.11 wireless mesh networks. To enhance thethroughput, they further present a new routing metric to select themost efficient route among all feasible ones. In [35], authors explorethe problem of admission control and scheduling of flows in multi-hop WiMAX networks. In particular, they propose efficient heuristicalgorithms for scheduling flows in a centrally scheduled multi-hopWiMAX network. Finally, in [37] authors propose an integrated ad-mission control and routing mechanism for multi-rate wireless meshnetworks. They discuss a DCSPT method for available bandwidthestimation, based on dual carrier sensing with parallel transmissionawareness and introduce a packet probing-based available bandwidthestimation method, suitable for legacy device implementations. Thesetechniques are further integrated in an admission control mechanismdesigned for a hop-by-hop routing protocol (LUNAR), enablingalternate route identification when shortest paths are congested.

IV. D ESIGN OFINTERFERENCE-AWARE ADMISSION CONTROL

ALGORITHM (IAC)

As emphasized before, the main objective of IAC is to mitigatethe overhead involved in estimating available bandwidth informationat each interfering neighbor seen in existing literatures such asCACP [12], MARIA [17]. For this purpose, IAC requires each nodelocally to estimate the channel utilization ratio and then compareswith the threshold to determine whether the current channel utiliza-tion has exceeded the minimum permissible threshold. If exceeded,each node is required to compute the available bandwidth andbroadcast this information to all the interfering nodes. IAC requiresall the receiving nodes to buffer these broadcasts and as a resultwhen a new flow request arrives at each node, it locally determineswhether the incoming flow can be accepted by comparing the newflow requirement with the buffered bandwidth information.

In this sequel, initially we discuss how to determine the local andneighbor available bandwidth of a given node. Then, we focus onthe design of the proposed algorithm, IAC and several challengesassociated with the design of IAC.

A. Estimation of local available bandwidth

As mentioned in section II, the local available bandwidth is definedas theunusedor spare bandwidth observed by a given node. Themost common way to measure the available bandwidth (ρavail) isto determine the channel utilization (υ). Given, ρ, the total channelbandwidth andυ, the channel utilization, the available bandwidth(ρavail) is defined as [26]:

ρavail = (1− υ)× ρ

Many techniques have been proposed for measuring the channelutilization as given in [12], [14], [20]. In this paper, we use thefraction of thebusy timeduring every time periodT as an indicationof the channel utilization at a node. A node can easily detect the stateof the channel asidle or busyby monitoring the network activities

4

0

10

20

30

40

50

60

70

80

90

100

0.9 1.5 2 2.5 2.9 3.5

Time -->

Ch

an

nel U

tilizati

on

(%

)

Node 2

Node 6

Node 14

Fig. 4. Channel utilization (%) with respect to three flows (1, 2, 3) vs Time.RTS/CTS mechanism is used.

since the IEEE 802.11 is a CSMA-based MAC protocol, workingon the virtual and physical carrier sensing mechanisms. The wirelessmedium around a node (the channel) is determined to bebusydue toany of the following reasons: (a) transmitting or receiving the packetsby the node; (b) sensing a signal with strength greater thancarrier-sense threshold. Nonetheless, the node cannot decode the contents ofthe message if it is within the carrier sensing range and outside thetransmission range of the sender; (c) the network allocation vector,(NAV), which indicates the channel is busy due to activity of othernodes [22].

We estimate the channel utilization,υ, as a fraction of the timethe channel is detected as busy due totransmissionor receptionofpackets orsensing(virtual or physical carrier sensing) other packettransmissions. The channel utilization,υ, and the local availablebandwidthρlocal

avail are expressed as follows:

υ =Tbusy

T

ρlocalavail = ρ× (1− υ) (1)

whereTbusy is the amount of busy channel time monitored by thenode during every time periodT .

Fig. 4 depicts the channel utilization observed by the nodes withrespect to the simulation topology in fig. 1. It can be observed fromthe figure that at time t =2.9 seconds, the channel utilization fornode 6 is 92% and for node 14 is 46%. As a result, when node 14admits the flow at time t =3 seconds, the channel utilization observedby node 6 reaches above 99% and therefore, QoS of flow 2 fallsbelow the desired level. From this observation, we can conclude thatchannel utilization provides an accurate estimation of local availablebandwidth,ρlocal

avail, at a given node.

B. Estimation of neighbor available bandwidth

As discussed in section II, an admission control algorithm shouldtake into consideration both thelocal available bandwidthandneighbor available bandwidthof a given node to maintain the qualityof all admitted flows in the network. However, estimation of a node’sown local available bandwidth cannot provide information aboutits neighbor available bandwidth or the local available bandwidthinformation of nodes in its transmission or carrier sensing range.For e.g., as mentioned in section II, at time t =3 seconds, node 14estimated its unused bandwidth as 53.5%. On the other hand, node 14does not know the amount of bandwidth available at node 6 which is7%. Therefore, the amount of bandwidth available for a given node(ρnode

avail) is theminimumof the local available bandwidth (ρlocalavail) and

neighbor available bandwidth (ρneigh

avail ). Consequently, we propose anactive technique to attain the local available bandwidth informationof the nodes within the transmission and carrier sense range of aparticular node. In active technique, nodes actively exchange theirlocal available bandwidth information between each other. Sincenodes cannot communicate directly with the nodes in their carriersense range, such information exchange is done with high transmit

power which in turn results in high overhead [12]. Nonetheless,these active techniques can provide accurate information about thelocal available bandwidth of the interfering neighbors. Therefore,to compensate for the overhead caused by the active technique,we propose an optimized version of the IAC algorithm known asInterference-Aware Admission Control Optimized(IAC/O) that usesa threshold based approach for the exchange of bandwidth relatedmessages.

In IAC/O, each mesh node monitors the state of the surroundingwireless medium to measure the channel utilization. We denote themeasured channel utilization asυc. It is important to note that thisprocess has little overhead since the IEEE 802.11 CSMA/CA protocolworks on virtual and physical carrier sensing [22], [13]. During everytime periodT , the node computes the fraction of busy channel time,Tbusy, to determine the channel utilization,υc, whereυc =

Tbusy

T.

The IAC/O is adual-threshold schemerelying on two thresholdsδυ

min and δυmax. δυ

min is the lower channel utilization thresholdwhereasδυ

max is the upper channel utilization threshold. The fol-lowing relationship exists between the thresholds and the maximumchannel utilization,0 < δυ

min < δυmax < υ. We limit the maximum

channel utilization of the network toδυmax to avoid congestion

due to over-utilization of the resources due to incorrect admissions.Therefore, when the channel utilization exceedsδυ

max, the nodes inthe interfering range set their available bandwidth tozeroand rejectnew flow requests2.

As depicted in fig. 5, given the two thresholdsδυmin andδυ

max, wedivide the channel utilization intothree ranges:rl (lower range),rm

(medium range) andrh (higher range). Once the channel utilizationυc, is determined, each node comparesυc with the thresholds,δυ

min

and δυmax, to determine the range. Based on the range in whichυc

lies, the node performs the following operations:

U=0 U=maxU= min U= max

rl rm rh

Fig. 5. Channel utilization ranges defined on the basis of thresholds,δυmin

and δυmax. When υc ∈ rm, the nodes are moderately congested whereas

whenυc ∈ rh, the nodes are completely congested.

• υc ∈ rh (completely congested): In this range, the follow-ing relationship exists between the thresholds and the currentchannel utilization,δυ

max ≤ υc ≤ υ. Assume that nodei’schannel utilization exceeded the upper thresholdδυ

max, duringtime periodT . Since the channel utilization exceeded the upperthreshold, nodei assigns theρlocal

avail to zero and broadcasts anevent message,bandwidth lost (bwlost), to all the nodes inits transmission and carrier sensing range. When the interferingneighbors of nodei receive thebw lost message, they decidetheir available bandwidth (ρnode

avail) as follows:

ρnodeavail(j) = min (ρlocal

avail(j), ρneigh

avail (i) = 0), ∀j ∈ N

whereN is the set of interfering neighbors of nodei. Since theinterfering nodes inN do not have enough available bandwidthto admit a new flow, the channel will not be overused andaccordingly, the quality of all admitted flows will be preserved.

2Note that, if IAC depend only onδυmax, then there is high probability for

over-utilization of the resources due to incorrect admissions of flows. Thusto avoid this situation, before reaching the maximum utilization IAC requiresthe nodes to give a notification message to its interfering neighors about theircurrent channel utilization. To achieve this objective, IAC requires a minimumthreshold,δυ

min.

5

• υc ∈ rm (moderately congested): In this range, the followingrelationship exists between the thresholds and the current chan-nel utilization,δυ

min ≤ υc < δυmax. Suppose that during time

periodT , nodei’s υc exceeded the lower threshold,δυmin. When

this occurs, nodei computes its available bandwidth,ρlocalavail, as

follows:

ρlocalavail(i) = ρ× (δυ

max − υc) (2)

and broadcasts an event message,bandwidth critical(bw critical), to all the nodes in its transmission and carriersensing range (i.e. interfering neighbors). When the interferingneighbors of nodei receive thebw critical message, theyrecord the local available bandwidth information of nodei anddetermines their available bandwidth as follows:

ρnodeavail(j) = min (ρlocal

avail(j), ρneigh

avail (k))

∀j ∈ N,∀k ∈M (3)

whereN and M are the set of interfering neighbors of nodesi and j respectively.ρneigh

avail (k) is the neighbor available band-width of all interfering nodes ofj obtained viabw critical orbw lost message.In order to make accurate decisions in this range, we require themoderately congested nodes to exchange information about theavailable bandwidth with their interfering nodes. This informa-tion exchange can be eitherperiodic exchangeor event-drivenexchange. In the periodic exchange, thebw critical message issent to the interfering neighbors everyτ second provided thatthe current channel utilization is within rangerm. This approachhas high overhead and therefore to compensate for the overheadinvolved in periodic message exchange, we employ the secondapproach of information exchange known asevent-driven. In thisapproach, whenever the channel utilization changes to a newvalue within the rangerm, a bw critical message is sent to theinterfering neighbors. This scheme is accurate and has loweroverhead. Our simulation results also demonstrate thatevent-driven exchange performs better when compared toperiodicexchange.

• υc ∈ rl (uncongested): In this range, the following relationshipexists between the thresholds and the current channel utilization,0 ≤ υc < δυ

min. Since the channel utilization is below thelower threshold, the nodes do not exchange any bandwidthrelated information with their interfering neighbors to reducethe overhead.

In addition to the two event messages (bw critical andbw lost), weemploy a third message known asbandwidth recover (bwrecover).Assume that the channel utilization,υc, exceededδυ

max or δυmin

during time periodT . In the next time period, ifυc falls below thelower threshold by∆ (i.e. υc < (δυ

min−∆)) a bw recovermessageis transmitted to the interfering neighbors. As a result,bw recovermessage prevents the system resources from being under-utilized.The threshold∆ is used to reduce the overhead associated with thevalue of υc oscillating between the ranges. A smaller∆ providesefficient channel utilization at the expense of overhead. On the otherhand, a larger∆ incurs low overhead but may result in under-utilization of channel resources. Therefore, while designing∆, weneed to carefully maintain a tradeoff between overhead and channelutilization.

It is important to note that the event messages are exchanged onlywhen the channel utilization exceeds the thresholdδυ

min and hence,IAC/O has less overhead compared to the active techniques discussedin [12]. Algorithm 1 summarizes the estimation of neighbor availablebandwidth.

Algorithm 1 IAC/O: Estimation ofρneigh

avail

# Initialize Stateto true.State= true;if (υc < δυ

min) thenif (State= true) then

do nothing;else

# State is equal to falseif ((υc + ∆) ≤ δυ

min) thenState= true;send bw recover();

end ifend if

elseif (υc ≥ δυ

min andυc < δυmax) then

State= false;ρlocal

avail = ρ× (δυmax − υc);

send bw critical (ρlocalavail);

elseif (υc ≥ δυ

max andυc ≤ υ) thenState= false;ρlocal

avail = 0;send bw lost (ρlocal

avail);end if

end ifend if

In the following sections, we present several challenges encoun-tered while implementing IAC/O algorithm and discuss solutions foreach.

C. Model for neighbor communication

In a CSMA-based MAC protocol [22], [13], the carrier sensingrange of wireless nodes is typically larger than their transmissionrange. Nodes that are outside the transmission range but within thecarrier sensing range can sense packet transmissions; however, theyare unable to interpret the contents of the packet. Therefore, directcommunication is ineffective for sharing bandwidth information withsuch nodes.

Several methods for communicating with carrier sensing neighborshave been discussed in [12], [14]. In IAC/O, we employ the hightransmit power approach presented in [12] for sharing bandwidth re-lated information. This technique relies on a high transmission powerto broadcast the bandwidth related event messages (bw critical,bw lost, bw recover) to the nodes within the transmission and carriersensing range. Using this high transmit power approach, all nodeswithin the interfering range of the sender node are contacted. It isnot hard to deduce that these high power messages may incur highinterference as opposed to data transmission with normal power.However, it is important to note that, this high transmit powerapproach does not induce much overhead in IAC/O due to thefollowing reasons:(a) Event messages are exchanged only whenthe channel utilization exceeds the threshold. For e.g., whenυc

exceedsδυmax, a bw lost message is sent to the interfering nodes

to inform them about the available bandwidth resources. Therefore,transmitting an event packet is an infrequent event as opposed tothe data transmission.(b) Event messages are broadcast packets.Therefore, a single message from the sender node will reach all thenodes within the transmission and carrier sensing range of the sender.

6

D. Model for Available Bandwidth Allocation

As mentioned in section IV.B, during every time periodT , eachnode computes the current channel utilizationυc and determineswhether υc has exceeded the thresholdsδυ

min and δυmax. If the

channel utilization has exceeded the thresholds, the node computesthe local available bandwidth and transmits the corresponding eventmessages with this information within its neighborhood using themechanism described in section IV.C. In this section, we discussthe issue of sharing the remaining bandwidth among the interferingneighbors of a given node so that the channel resources are efficientlyutilized.

A BGC

flow1 flow2 flow3 flow4

550m

200 200 250

Fig. 6. A simple topology. Nodes A, B and G are interfering neighbors ofnode C. Nodes A and G are the direct neighbors of node C; node B is thecarrier sensing neighbor of node C.

For example, consider the topology in fig. 6 and the channelutilization observed by node C in fig. 7. Fig. 7(a) depicts theutilization of channel at time t-1, when no flows are admitted inthe network. Suppose that at time t, three flows are admitted bynodes A, B and C. Fig 7(b), depicts the current channel utilizationobserved by node C at time t. However, after the initiation ofthree flows, the utilizationυc exceeded the lower thresholdδυ

min.Consequently, node C computes the local available bandwidth usingeqn. (2) and broadcasts the event message,bw critical, to nodeswithin its neighborhood. When the interfering neighbors of node C(nodes A, B, and G ) receive thebw critical message, each nodedetermines the available bandwidth as given by eqn. (3). How shouldnodes A, B, C and G share the remaining bandwidth at node C so thatthe quality of all the admitted flows are preserved? It is important tonote that we should take into account the issue of bandwidth sharingonly whenυc is in the rangerm. Whenυc ∈ rh, the local availablebandwidth is set as zero and hence, bandwidth sharing need not beconsidered.

As discussed in previous sections, since the interfering neighborsof a given node cannot communicate directly, the sharing of availablebandwidth is NOT a straight forward process. Additionally, since thegiven node is unaware of the number of carrier sensing neighbors, afair sharing of available bandwidth is not possible. In this section, wediscuss two unfair bandwidth allocation schemes:First Come Firstserve (FCFS) allocationand Maximum Channel Utilization based(MCU) allocation.

We present each of the following allocation schemes with respectto the topology in fig. 6.

1) First Come First Serve (FCFS) Allocation:In this approach,any interfering neighbor of a given node can admit the flow if itsavailable resources,ρnode

avail, satisfies the bandwidth requirement of theflow. As an example, suppose that at time t+1, node G receives a newflow request and determines whetherρnode

avail satisfies the bandwidthrequirement of the new flow. If enough resources are available, nodeG will accept the new flow. Fig. 7(c) shows the channel utilizationobserved by node C after the initiation of flow by node G. However,it is evident that there is no fixed pattern of sharing the availablebandwidth between one or more flows. As discussed in previoussections, since the carrier sensing neighbors may not be able to

t-1 t

maxmin

Free Channel R

t

A R

t+1

min

B C Free

max

(a) Channel Utilization at time (t-1,t)

(b) Channel Utilization at time (t,t+1)

after admitting flows by Node A ,B, and C

t+1

A R

t+2

B C G

max

(c) FCFS Allocation scheme at time (t+1, t+2):

Flow from Node G admitted

min

t+1

A R

t+2

B C

max

(d) Utilization based Allocation scheme at time

(t+1, t+2): Flow from G not accepted.

min

A|B|C

T

Fig. 7. Channel Utilization observed by node C during every time periodT. The channel bandwidthρ(υ − δυ

max) is reserved (R) to avoid channelover-utilization.

communicate with each other, there is a high probability that one ormore flows are initiated simultaneously which can cause starvationto the existing flows.

To compensate for this drawback of FCFS approach, IAC/O designemploys δυ

max in dual-threshold scheme. We have discussed insection IV.B that IAC/O requires each node to limit the channel uti-lization toδυ

max and the remaining channel bandwidthρ×(υ−δυmax)

is reserved. Thus, even though one or more flows are admittedconcurrently, tuningδυ

max approximately can avoid channel over-utilization.

2) Maximum Channel Utilization Based Allocation (MCU):In thisapproach, each node gets access to the remaining channel bandwidthbased on the channel utilization history. In other words,“The more the node utilizes the channel, the more the share of theremaining available bandwidth”.Fig. 7(b) shows the state of the channel observed by node C after theinitiation of flows from node A, B and C. Hence, after the admissionof flows from A, B and C, the available channel bandwidth is reducedto ρ × (δυ

max − υc). For example, consider that during next timeperiod t+1, each node (A, B, C and G) receives a new flow requestfrom its corresponding application layer. In this scenario, we needto determine how efficiently the remaining channel resources canbe shared among nodes A, B, C and G so that the channel is notoverused.

Firstly, according tomaximum channel utilization based allocation(MCU) scheme, node G will not receive a share of the remaining

7

bandwidth since it has not utilized any fraction of the channelresources in previous time slots. As a result, node G rejects the newflow request.

Secondly, we have to determine how to allocate the remainingbandwidth resources among nodes A, B, and C. Since node C isunaware of the number of active carrier sensing neighbors, a fairshare of bandwidth is not possible. However, each node is aware ofthe fraction of the bandwidth resources it used during the previoustime periodT . As an example, suppose that each flow from nodesA, B, and C usedρused share of the total channel bandwidthρ. Eachnode computes the fraction of bandwidth consumption (x) as follows:

x =ρused

ρ

Oncex is determined, each node individually calculates the remainingbandwidth share (ρnode

avail) that can be consumed without affectingthe quality of existing flows. Supposei is the node that transmittedbw critical message to its interfering range and theρnode

avail of eachinterfering node in this approach is expressed as follows:

ρnodeavail(j) = min (ρlocal

avail(j), x× ρneigh

avail (i), ρneigh

avail (k))

∀j ∈ N,∀k ∈M − i (4)

whereN is the set of interfering neighbors of nodei andM is the setof interfering neighbors of nodej. Fig. 7(d) shows the MCU scheme.

The upside of the scheme is that: (a) The remaining bandwidthwill not be subjected to an excessive amount of demand. Therefore,channel will never be over-utilized; (b) Nodes that always have datato send are never deprived of an opportunity to do so. However, thedownside of the scheme is that heavily loaded nodes (nodes A, B,and C) mayhog the remaining bandwidth and the new nodes (fore.g., node G in fig. 6) are never given an opportunity to transmit.

It is important to note that the upside of both schemes rely on thedifference betweenδυ

max andδυmin. The smaller the difference, MCU

based allocation performs better because the remaining bandwidth canbe efficiently utilized without subjecting to an excessive amount ofdemand. However, the larger the difference, FCFS based allocationscheme performs better because of the following reasons: (i) Allnodes get a chance to transmit as opposed to MCU based scheme;(ii) Even though one or more flows are admitted simultaneously, theover-utilization of channel can still be avoided.

V. A LGORITHM IMPLEMENTATION

In this section, we present the implementation of IAC/O algorithmin the context of the AODV [15] routing protocol. However, theconcepts under discussion can be applied to other multihop routingprotocols [21] as well. Our main contributions in this section are:

• Computation of bandwidth requirement of a flow.

• Estimation of intra-flow interference.

• Implementation of IAC/O.

A. Computation of bandwidth requirement of a flow (ρreq)

In this section, we present the scheme for computation of thebandwidth requirement of a new flow so that the admitting node candetermine whether its available bandwidth can support this new flow.Initially, the data rate of the application must be converted into thecorresponding channel bandwidth requirement. In this process, themedium access control and network layer overhead should be takeninto account. For IEEE 802.11 MAC using RTS/CTS/DATA/ACKhandshake [22], [13], the transmission time of each data packet,Tdata, can be determined as follows:

Tdata = Trts + Tcts +S

ρ+ Tack + Tdifs + 3Tsifs (5)

where,S is the size of the data packet including the MAC and IPheader length,Trts, Tcts andTack represent the time for transmittingRTS, CTS, and ACK packets, respectively.Tsifs and Tdifs denotethe inter-frame spaces SIFS and DIFS respectively.ρ is the channelbandwidth.

If the application at the sending node generatesN packets persecond, then its corresponding bandwidth requirement (ρreq) can beexpressed as follows:

ρreq = N × Tdata × ρ (6)

B. Estimation of intra-flow Interference

In multihop networks, we should take into account the interferencecaused when the nodes on the path of same flow contend with eachother for the channel bandwidth, known as intra-flow interference [3].To understand the significance of intra-flow interference, consider the

A B C D E F250 250250 250

Destinationsource

Fig. 8. Example 1:A is the source and F is the destination. The carriersensing range of node A is shown as shaded area. In this topology all nodesare equally spaced at 250m apart.

multihop path in fig. 8. Here all nodes are equally spaced at 250mapart. The shaded area represents node A’s carrier sensing rangeandthe dashed circles represent each node’s transmission range. In fig.8, node B is the direct neighbor and node C is the carrier sensingneighbor of node A. Thus, nodes A, B and C cannot transmit a flowat the same time. Hence, the number of nodes that interfere with eachnode’s transmission in a multihop path is called ascontention countand denoted asNc. In fig. 8, the contention count at node A is 2.Therefore, the bandwidth consumption of the flow,ρc, at each nodei can be expressed as follows [12]:

ρc(i) = Nic × ρreq (7)

Generally, the computation of contention count is not a straightforward process. Some of the earlier works [27], [29] computedNc by taking into consideration that the carrier sensing range isapproximately twice the size of the transmission range and hence, thenumber of two-hop neighbors on the path can be used to approximateNc. However, we show that this approximation can not be appliedto all cases. For example, consider the topology shown in fig. 9. Wemodify the transmission distance between each node so that node A’scarrier sensing range now includes nodes B, C, D and E. In this case,the contention count at node A is 4 which is clearly higher than thenumber of two-hop neighbors.Note: We do not include destination in the contention count com-putation, because the destination node only passively receives trafficand hence does not contribute to channel contention. For example,in fig. 9, the contention count at node C is only 4 (Nodes A, B, Dand E).

We can estimateNc in a multihop environment, if some locationinformation of nodes could be distributed in the network during routediscovery phase. For example, each node may append its locationinformation in the route discovery packets before forwarding them.Given the location information of each node in the packet,Nc canbe determined as given in algorithm 2.

8

A B C D E F200 60 20060 200

Destinationsource

Fig. 9. Example 2: A is the source node and F is the destination node.The carrier sensing range of node A is shown as shaded area. Here, thetransmission distance between the nodes AB, BC, CD, DE and EF aremodified as 60m, 200m, 60m, 200m and 200m respectively.

Algorithm 2 IAC/O: Estimation ofNc of each node

# N ic ← contention count of each Nodei;

# Snode ← set of nodes that marked the packet with location;# CSR← carrier sensing range;# ID(j)← ID of nodej;# Dist(i, j)← Distance between nodei andj;N i

c ← 0for all j ∈ Snode do

if Dist(i, j) ≤ CSR & ID(j) 6= DestinationthenN i

c ← N ic + 1

elsedo nothing;

end ifend for

The drawback of this approach is that the length of the messageincreases with the number of hops which in turn increases the packettransmission time. However, it is important to note that the routediscovery process is initiated only when a node has data to send,which is a much rarer event as opposed to the data transmissionitself.

C. Implementation of IAC/O

In previous sections, we discussed on estimating available band-width at each node by taking into account the effect of inter-flowand intra-flow interference. In this section, we focus on setting up abandwidth aware end-to-end route between a source and destinationpair. In other words, the proposed approach ensures that each node onthe selected path has enough available bandwidth to support the newflow from source node. We discuss the algorithm implementation inthe context of AODV routing protocol.

1) Route Discovery in AODV:In AODV, when a source has datato send to an unknown destination, it broadcasts a Route Request(RREQ) for that destination [15], [3]. At each intermediate node,when a RREQ is received, a route to the source is created andstored. Each receiving node rebroadcasts the RREQ, if: (a) It hasnot received this RREQ before; (b) It is not the destination; (c) Itdoes not have a current route to the destination. If the receiving nodeis the destination or has a fresh route to the destination, it generates aRoute Reply (RREP). The RREP is unicast in a hop-by-hop fashionto the source. As the RREP propagates, each intermediate node alsorecords the route to the destination. When the source receives theRREP, it records the route to the destination and can begin sendingdata. If multiple RREPs are received by the source node, the routewith the smallest metric is chosen.

2) QoS-aware Route Discovery:To facilitate QoS provisioning,we add the following fields to RREQ and RREP packets of the AODVprotocol:

• ρreq: bandwidth requirement of the new flow as calculated bythe source node.

• location list: each node appends its location in the route discov-ery packets.

When the source node receives a new flow request, it broadcaststhe RREQ message. Besides the original fields in RREQ, the routerequest now also contains the bandwidth requirement of the new flow(ρreq) and the location of the source node. When each intermediatenode receives the route request, it performs admission control bydetermining whether the node has enough resources (ρnode

avail) tosupport the new flow. Note that, since the nodes do not have anyinformation about the contention count, they compare the requiredbandwidth of the flow,ρreq, with the available resources,ρnode

avail. Ifthe bandwidth request is smaller than the available resources, thereceiving node appends its location information and rebroadcasts therequest. Otherwise, the node will drop the request. As the RREQpropagates, each node performs similar operations.

When the destination node receives RREQ, it sends a route replyback to the source along that route. Besides the original fields ofRREP, the reply message contains the bandwidth requirement of thenew flow and the location id of all the nodes in the path betweensource and destination. As mentioned before, the location of the nodesis used to determine the contention count of each node in a multihoppath (see algorithm 2). As the RREP propagates, each node alongthe path computes the contention count using the locationlist anddetermines whether the available bandwidth of the node,ρnode

avail, cansatisfy the bandwidth consumption of the flow,ρc. If there are enoughresources at the node, the RREP is forwarded to the next hop. At thesame time, a soft reservation for bandwidth is set up at the node whenRREP propagates back to the source node. On the other hand, if thereare not enough resources at the node, afailure message is forwardedalong the path towards the destination. Thus, the soft reservationsof bandwidth along the route are explicitly taken apart when nodesreceives the failure message. Finally, when the RREP arrives at thesource node, enough end-to-end bandwidth has been reserved andthe flow can start. Thus, admission control is completed only whenRREP propagates back to the source node. Note that unlike in AODV,we require that only the destination node can generate an RREP andnot the intermediate nodes. Hence, when the source node receivesthe RREP, it can ensure that enough resources are allocated end-to-end between the source and destination nodes. Moreover, due to thechange of wireless channel condition or network topology, if any nodealong the established route(s) detects that QoS requirements can notbe maintained, that node must send an QoSLost message as seen inAODV-QoS[33] to inform the source nodes about the unmaintanableflows. The corresponding source nodes have to reinitiate new flowrequests.

Since each node has the information of the available bandwidthat its interfering neighbors, we can clearly see that the admissioncontrol algorithm admits a flow without starving the existing flowsin the neighborhood.

VI. A NALYSIS OF DUAL-THRESHOLD SCHEME

In this section, we discuss two variants of dual-threshold schemeknown asdeterministicand adaptivedual-threshold scheme. Sincethe lower channel utilization threshold has a significant role inIAC/O, in both schemes, we setδυ

max to a pre-determined valueand focus on the estimation ofδυ

min. In this paper, we tuneδυmax

to 90%. This is because it is shown in [30] that when the network

9

utilization exceeds 95% (with RTS/CTS enabled) or 90% (withoutRTS/CTS) the throughput drops quickly and the delay variationincrease dramatically. Clearly, this point is the optimal operating pointthat we should tune the network to work around, where the throughputis maximized and the delay and delay variation are small.

A. Deterministic dual-threshold scheme

In deterministic dual-threshold scheme, we set both the thresholdsδυ

min and δυmax to a pre-determined value. Although, it is easier to

calculate the thresholds in this scheme, it is not adaptable to thenetwork state.

B. Adaptive dual-threshold scheme

The main goal of the adaptive scheme is to use historical data toidentify trends in channel utilization and predict future utilization ina reasonably accurate fashion. Based on the estimated future channelutilization, we adapt the lower threshold,δυ

min. We employ akth-step linear predictorto estimate the future channel utilization [31].We define the following notations used in estimation:

• υ[n]: Channel utilization at timen.

• p: Number of past measurements used for prediction

The problem can be formulated as a linear prediction and expressedas follows:

υ̂[n + k] =

p−1∑

a=0

υ[n− a]wn[a] = WTn V [n] (8)

In eq. (8), the right side are the past samples and the predictioncoefficients,W T

n and the left side represents the predicted values. Theprediction coefficients are time varying and updated by minimizingthe mean square errorζ where

ζ = E[e2[n]] (9)

V [n], Wn, ande[n] are defined as follows:

V [n] = [υ[n], υ[n− 1], ...., υ[n− p + 1]]T (10)

Wn = [wn[0], wn[1], ...., wn[p− 1]]T (11)

e[n] = υ[n + k]− υ̂[n + k] (12)

On the basis ofe[n], the prediction coefficients are updated to reducethe error. The coefficient update relation is expressed as follows:

Wn+1 = Wn + µe[n]V [n] (13)

whereµ is the constant step size [31] that controls the amount ofinformation used to update each coefficient.

Estimation of δυmin: During every time periodT , each node

computes the current channel utilization and determines the futurechannel utilization,̂υ, as given by eq. (8). Once the future utilizationis predicted, our algorithm determines whetherυ̂ exceeds the upperutilization threshold,δυ

max. If the predicted utilization is greater thanthe upper threshold, the node may send abw critical message to allthe nodes within its contention range with the available bandwidthset toρ×δυ

max−υc. Algorithm 3 summarizes the estimation ofδυmin.

VII. PERFORMANCEEVALUATION

In this section, we evaluate the performance of IAC/O in termsof throughput, delay and packet loss rate for VoIP calls. Throughsimulations, we demonstrate that IAC/O maintains the quality of allaccepted traffic flows in the network.

Algorithm 3 IAC/O: Adaptive estimation ofδυmin

# Assignδυmax to a pre-determined value;

Given p, W Tn , estimate the future channel utilization

(υ̂) as given in equation(8);if υ̂ > δυ

max thenδυ

min = υc

#send bwcritical message.else

do nothing;end if

A. Simulation Parameters

We use ns2(v2.29) [16] to evaluate the performance of IAC/O. Thenetwork topology consists of 25 mesh nodes randomly located in a1500m X 1500m area. In the simulations, traffic sources are modeledas CBR flows. We consider an ITU-T G.711 codec packetizing 10msof audio data in 80 bytes of payload (data rate of 64Kbps) [7]. TheMAC layer protocol is IEEE 802.11 with 250m radio transmissionrange and 550m carrier sensing range. The bandwidth of the wirelesschannel is set to 2Mb/s. In Table I, we provide the IEEE 802.11system parameters used for our simulation.

TABLE IIEEE 802.11 SYSTEM PARAMETERS

Slot time 20µsSIFS 10µsDIFS 50µs

Phy Header 192 bitsMAC Header 224 bits

PLCP Data Rate 1Mb/s

Following the work in [30], we tune the upper utilization thresholdδυ

max to 90% and 10% for∆. We also used a time period of 500msto monitor the channel utilization.

B. Performance Results

1) Throughput of flows:In this scenario, we established six calls(or flows) between random pairs of source and destination with aninter-arrival time of 10s. Fig. 10 depicts the throughput of the sixcalls in the absence of admission control. The graph illustrates thatas the simulation progresses and more calls are accepted, the channelis overused and the throughput of the calls (flows 2 and 4) startsdegrading. Two observations can be made from fig. 10: (a) At timet =40 seconds, when the sender node admits flow 4, the throughputof the flow 2 degrades by 33%; (b) At time t =50 seconds, when thesender node admits the flow request 5, the throughput of the flow 4degrades by 30%. This is because the nodes along the path of flow2 and flow 4 do not have enough resources to support the new flowrequests 4 and 5 respectively and hence, the contention from theseflows degraded the throughput of flows 2 and 4. Fig. 11 shows thethroughput of the calls in the presence of IAC/O. Two observationscan be deduced: (a) The number of calls admitted in the network isreduced to five; (b) The throughput for all the calls are maintainedat a constant value. This is because at time t =40 seconds, IAC/Orejects call request 4, when it finds that the neighbors of call 4 donot have enough resources to support the call.

2) Delay and Packet Loss Rate of flows:In this scenario, wedemonstrate the delay of calls both in the absence and presence ofadmission control algorithm. From fig. 12, we can see that after timet =50 seconds, the delay of calls 2 and 4 is increased sharply to 0.8s

10

10 20 30 40 50 60 70 80 900

10

20

30

40

50

60

70

80

90

Time (Seconds)

Thr

ough

put (

kbps

)

Flow 1Flow 2Flow 3Flow 4Flow 5Flow 6

Fig. 10. Throughput of six flowsvs Time when no admission controlalgorithm is employed. RTS/CTS mechanism is used.

10 20 30 40 50 60 70 80 900

10

20

30

40

50

60

70

80

Time(Seconds)

Thr

ough

put(

kbps

)

Flow 1Flow 2Flow 3Flow 5Flow 6

Fig. 11. Throughput of five flowsvs Time when IAC/O is employed.RTS/CTS mechanism is used.

and 1.4 seconds respectively. As discussed before, since the delayfor an acceptable VoIP flow is less than 200ms according to ETSIrecommendation, the delay experienced by calls 2 and 4 can havean adverse effect on the quality of calls. In contrast to the poor

0

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

10 20 30 40 50 60 70 80 90 100

Del

ay (

seco

nds)

Time (seconds)

Flow 1Flow 2Flow 3Flow 4Flow 5Flow 6

Fig. 12. Delay of six flowsvs Time when no admission control algorithmis employed. RTS/CTS mechanism is used.

performance in the absence of admission control, IAC/O performsadmission control to provide better service. Fig. 11 shows the delayof calls 1 to 3 and 5 to 6. We can observe that the delay is extremelysmall (less than 0.1s) for all the admitted calls. The reason for thisimprovement is similar to that explained in previous section. TableII shows the packet loss rate of each call both in the absence andpresence of admission control algorithm. It can be observed fromTable II that the packet loss rate is 0% for all admitted calls inthe presence of IAC/O. The smaller delay, lower packet loss rate andconstant throughput demonstrate that IAC/O can be used for networksto maintain real-time traffic applications, such as VoIP or VoD.

3) Periodic and Event Driven Message Exchange:In this scenario,we compare the performance of periodic and event-driven exchangeswith the case where “no exchange” algorithm is employed. The timeinterval (τ ) between two consecutive exchanges in periodic scheme istuned to 1 second. From Table III, two observations can be inferred:

0

0.2

0.4

0.6

0.8

1

1.2

1.4

10 20 30 40 50 60 70 80 90 100

Del

ay (

seco

nds)

Time (seconds)

Flow 1Flow 2Flow 3Flow 5Flow 6

Fig. 13. Delay of six flowsvs Time when IAC/O is employed. RTS/CTSmechanism is used.

TABLE IIPACKET LOSSRATE (%) FOR EACH FLOW

Call/Flow # w/o Admission Control IAC/O1 0% 0%2 21% 0%3 0.2% 0%4 53% N/A5 0% 0%6 0% 0%

(a) As compared to the case where “no exchange” is employed,both periodic and event-driven exchanges performed better. In “noexchange” scenario, the nodes do not transmit any bandwidth relatedinformation within their transmission and carrier sensing range andas a result, the nodes will admit the new flow requests which inturn affects the quality of the existing flows. (b) The performanceof the periodic exchange is decreased when compared to the event-driven exchange. This is because in periodic exchange schemenodes exchange their flow information periodically using high powerat every 1 second. Consequently these periodic exchanges usingextended power cause high overhead in the network and consumea part of the bandwidth which is available for data transmission.Clearly, this also implies that IAC/O can outperform existing schemessuch as MARIA that rely on high power periodic HELLO messagesto convey the bandwidth information in the neighborhood.

TABLE IIIPACKET DELIVERY RATE (%) FOR PERIODIC AND EVENT DRIVEN

MESSAGEEXCHANGES

Information Exchange Schemes Packet Delivery Rate(%)No Exchange 82%

Periodic Exchange 90.2%Event-Driven Exchange 99.3%

4) Sensitivity of Allocation Mechanisms to Thresholds:In thissection, we show the sensitivity of two allocation schemes FCFSand MCU to varying values of thresholds. As mentioned before, wereserve 10% of the maximum channel utilization to avoid congestionor temporary flutuations. Hence, the upper utilization threshold is setto 90%.

Fig. 14, 15 depicts the performance of FCFS and MCU basedallocation schemes for varying values of lower utilization threshold,δυ

min. It can be observed that whenδυmin is set to 50%, FCFS

performed better by admitting five flows as opposed to MCU thatadmitted only four flows. Since the bwcritical message is exchangedat δυ

min= 50%, MCU rejects flow requests 3 and 4 and reservesthe remaining bandwidth for flows 1 and 2. However, whenδυ

min

is set to 80%, both schemes showed poor performance. This isbecause the large value of the threshold prevented the congestednode from sending the event message to its neighboring nodes and

11

hence, IAC/O started the new call 4 at time t =40 seconds whichin turn degraded the quality of existing flows. Fig. 15 shows the

0

10

20

30

40

50

60

70

80

90

Flo

w 1

Flo

w 2

Flo

w 3

Flo

w 4

Flo

w 5

Flo

w 6

Flow Number

Avera

ge T

hro

ug

hp

ut

(kb

ps)

FCFS MCU

0

10

20

30

40

50

60

70

80

90

Flo

w 1

Flo

w 2

Flo

w 3

Flo

w 4

Flo

w 5

Flo

w 6

Flow Number

Avera

ge T

hro

ug

hp

ut

(kb

ps)

FCFS MCU

Fig. 14. Average throughput of each flow for deterministicδυmin= 50%,

80% under FCFS and MCU based allocation.

performance of allocation schemes forδυmin= 70%. We can see

that both schemes performed better in this scenario. In fig. 16, we

0

10

20

30

40

50

60

70

80

90

Flow1

Flow2

Flow3

Flow4

Flow5

Flow6

Flow Number

Avera

ge T

hro

ug

hp

ut

(kb

ps)

FCFS MCU

Fig. 15. Average throughput of each flow forδυmin= 70% under FCFS and

MCU based allocation.

present the performance of allocation schemes when adaptiveδυmin

is used. We employed akth-step linear predictor to estimateδυmin.

For different values of k, fig. 16 depicts the average throughput ofFCFS and MCU. It can be observed that when k= 1, both FCFSand MCU schemes performed better. However, when k= 2, FCFSshowed better performance as opposed to MCU. This is because aftertime t =20 seconds, the estimated future utilization exceededδυ

max

and hence, the congested node broadcasted abw critical messagewith the available bandwidth information (ρ(δυ

max − υc)) to all thenodes within its contention range. Since the MCU scheme reservesthe remaining bandwidth based on channel utilization, it rejects newflow requests (flows 3 and 4) and allocates the remaining bandwidthresources to flows 1 and 2. Even though the deterministic thresholdis simpler to compute, a series of simulations is required to tune thethreshold to a proper value. Therefore, for better sharing of bandwidthresources, we suggest to employ 1-step predictor for MCU and k-step(k > 2) for FCFS scheme.

VIII. D ISCUSSIONS ANDFUTURE WORK

In this section, we shed some light on the direction of our futureresearch by discussing several enhancements that can be made to theproposed algorithm.

0

10

20

30

40

50

60

70

80

90

Flow 1

Flow 2

Flow 3

Flow

4

Flow

5

Flow

6

Flow Number

Avera

ge T

hro

ug

hp

ut(

kb

ps)

FCFS, k=1

FCFS, k=2

MCU, k=1

MCU, k=2

Fig. 16. Average throughput of each flow for adaptiveδυmin under FCFS

and maximum utilization based allocation.

A. Impact of Mobility

In WMN, mobile clients rely on static mesh routers to get accessto the network. When a mobile client moves from a mesh accesspoint to another, it switches its connectivity to the closest accesspoint (called as handoff). Since the handoff is not transparent to thenomadic client, the new point of connection should be able to deliverthe required services to the client without any interruption. However,when the access point accepts the traffic from a mobile client, itshould also guarantee that there are enough resources to support thenew flow and ensure that the existing flows are not affected. Onepossible approach is to allocate a share of the bandwidth for thetraffic from mobile clients. We can easily incorporate the proposedscheme into IAC/O by intelligently tuning the two thresholds,δυ

min

andδυmax. Our future work is to investigate the performance of IAC/O

algorithm in the presence of traffic flows from mobile clients.

B. Multi-channel Mesh network

Interference among wireless links have a serious impact on theperformance of multi-hop wireless networks and hence, very fewflows can be admitted by the admission control algorithm to avoid theover-utilization of the shared channel resources. However, prior works[2], [3], [32] have shown that using multiple channels in the networkcan greatly improve the throughput by reducing the inter-flow andintra-flow interference. For further work, we also plan to enhancethe proposed algorithm by employingmultichannel, multiradioin thenetwork.

C. Flow Prioritization

At present, our algorithm is designed in such a way that all kindsof traffic are admitted with equal priority. In contrast tobest-efforttraffic, real-time flows have stringent QoS requirements; hence, werequire the admission controller to prioritize thereal-time trafficahead of thebest-effort traffic, if both flows arrive at the sametime. In other words, a priority should be assigned for each kind oftraffic (for e.g., high priority forreal-time traffic and low priority forbest-efforttraffic). For future work, we plan to extend the proposedalgorithm by incorporating QoS prioritization. For e.g.,real-timetraffic has stringent QoS requirements when compared tobest-efforttraffic and therefore, the admission control algorithm should assignhigher priority for real-time flows and lower priority forbest-efforttraffic.

IX. CONCLUSIONS

In this paper, we present IAC, an interference aware admissioncontrol algorithm for use in wireless mesh networks. The core concept

12

of IAC is to use a low overhead dual-threshold based approach toshare the bandwidth information with the neighbors in the interferingrange. As a result, IAC guarantees that the shared wireless bandwidthis not over-utilized and the quality of all existing flows are preserved.Moreover, IAC takes into account the intra-flow interference effectto estimate the bandwidth consumption of the flow in a multihoppath. We have also proposed two approaches of bandwidth allocation,FCFS and MCU, and demonstrated that proper tuning of thresholdscan lead to high performance of both schemes. Simulation resultsillustrate that IAC effectively limits the over-utilization of channelresources which in turn results in high throughput, low delay and lowpacket loss rate for all admitted flows. In this work, we assume error-free wireless channel and plan to extend IAC to perform admissioncontrol in the presence of channel errors.

REFERENCES

[1] I.F. Akyildiz and X. Wang, “A survey on wireless mesh networks,” inIEEE communication Magazine, vol. 43, no. 9, pp. s23-s30, Sept. 2005.

[2] A. Raniwala and T. Chiueh, “Architecture and algorithms for an IEEE802.11 based multi-channel wireless mesh network,” inProceedings ofInfocom, vol. 3, pp. 2223- 2234, Miami, FL, Mar. 2005.

[3] D. M. Shila and T. Anjali, “Load-aware Traffic Engineering for MeshNetworks,” inComputer Communications, vol. 31, no. 7, pp. 1460-1469,2008.

[4] Nortel[online], “http://www.nortel.com.”[5] Tropos Networks[online], “http://www.tropos.com.”[6] BelAir Networks[online], “http://www.belairnetworks.com.”[7] S. Garg and M. Kappes, “Can I add a VoIP call?,” inIEEE International

Conference on Communications, vol. 2, pp. 779-783, May. 2003.[8] K. Medepalli, P. Gopalakrishnana, D. Famolari and T. Kodama, “Voice

Capacity of IEEE 802.11b, 802.11a, 802.11g Wireless WLANs,”inProc. of IEEE Globecom 2004, vol. 3, pp. 1549-1553, Nov 2004.

[9] A. Zahedi and K. Pahlavan, “Capacity of a Wireless LAN with Voiceand Data Services,” inIEEE Transactions on Communications, vol. 48,Issue 7, pp. 1160 1170, Jul. 2000.

[10] X. Wang, A. Patil and W. Wang, “VoIP over wireless mesh networks:challenges and approaches,” inProceedings of Wireless Internet Con-ference (WICON 2006), Boston, MA, Aug. 2006.

[11] L. L. Peterson and B. S. Davie, “Computer Networks: A SystemApproach,” Morgan Kaufmann Publisher, 3rd edition, Oct. 1999

[12] Y. Yang and R.Kravets, “Contention-Aware Admission Control for AdHoc Networks,” inIEEE Transactions on Mobile Computing, vol. 4, no.4, July. 2005.

[13] IEEE Standard 802.11-1999, Part 11: Wireless LAN Medium AccessControl (MAC) and Physical Layer (PHY) Specifications, IEEE, 1999.

[14] I.D. Chakeres, and E.M Belding-Royer., “PAC: perceptive admissioncontrol for mobile wireless networks,” inQuality of Service in Hetero-geneous Wired/Wireless Networks, pp. 18-26, Oct. 2004.

[15] I.D. Chakeres and E.M. Belding-Royer, “AODV Routing ProtocolImplementation Design,” inProceedings of the International Workshopon Wireless Ad Hoc Networking (WWAN), Tokyo, Japan, Mar. 2004.

[16] K. Fall and K. Varadhan, NS notes and documentation, The VINTProject, UC Berkely, LBL, USC/ISI, and Xerox PARC, 1997.

[17] X. Cheng, P. Mohapatra, S. Lee and S. Banerjee, “MARIA: Interference-Aware Admission Control and QoS Routing in Wireless Mesh Net-works,” in IEEE International Conference on Communications, pp. 2865- 2870, May. 2008.

[18] C. H. Liu, A. Gkelias, and K. K. Leung, “Connection Admission Controland Grade of Service for QoS Routing in Wireless Mesh Networks,” inIEEE PIMRC 2008, pp. 1-5, 15-18 Sept., Canne, France.

[19] H. Zhu, V. O. K. Li, Z. Ma, and M. Zhao, “Statistical connectionadmission control framework based on achievable capacity estimation,”in IEEE International Conference on Communications, vol. 2, June 2006,pp. 748753.

[20] L. Chen and W. B. Heinzelman, “QoS-aware routing based onbandwidthestimation for mobile ad hoc networks,” inIEEE Journal on selectedareas in communications, vol. 23, no. 3, Mar. 2005.

[21] D.B.Johnson and D.A.maltz, “Dynamic Source Routing in AdHocWireless Networks,” inMobile Computing, 1996.

[22] G. Bianchi, “Performance Analysis of the IEEE 802.11 Distributed Co-ordination Function,” inIEEE Journal Selected Areas in Communication,vol. 18, no. 3, 2000

[23] S. Lee, G. Ahn, X. Zhang and A. T. Campbell, “INSIGNIA: An IP-Based Quality of Service Framework for Mobile Ad Hoc Networks,” inJournal of Parallel and Distributed Computing, vol. 60, pp. 374-406,2000.

[24] H. Wei, K. Kim, A. Kashyap and S. Ganguly, “On Admission of VoIPCalls Over Wireless Mesh Network,” inIEEE International Conferenceon Communications, vol. 5, pp. 1990-1995, June. 2006.

[25] G. Ahn, A-T. Campbell, A. Veres and L. Sun, “SWAN: servicedifferentiation in stateless wireless ad hoc networks,” inProceedingsof Infocom, vol. 2, pp. 457- 466, 2002.

[26] R. Prasad, C. Dovrolis, M. Murray, and K. Claffy, “Bandwidth estima-tion: metrics, measurement techniques, and tools,” inIEEE Network,vol. 17, no. 6, pp. 27-35, Nov. 2003.

[27] C. Rodrigo Cerveira and L.H. M. K. Costa, “A Time-based AdmissionControl Mechanism for IEEE 802.11 Ad Hoc Networks,” inSpringerBoston, vol. 211, 2006.

[28] D. Niculescu, S. Ganguly, K. Kim, and R. Izmailov, “Performance ofVoIP in a 802.11 Wireless Mesh Network ,” inProceedings of Infocom,Barcelona, Spain, pp. 1-11, Apr. 2006.

[29] R. de Renesse, M. Ghassemian and V. Friderikos “AdaptiveAdmissionControl for Ad Hoc and Sensor Networks Providing Quality of Service,”Technical Report, Center for Telecommunications Research, King’sCollege London, UK, May. 2005.

[30] H. Zhai, X. Chen and Y. Fang “How well can the IEEE 802.11 wirelessLAN support quality of service?,” IEEE Transactions on WirelessCommunications, vol. 4, no.6, pp.3084 - 3094, Nov. 2005.

[31] S. Haykin “Adaptive Filter Theory. (3rd edition),”Prentice Hall, 1996.[32] J. Tang, G. Xue and W. Zhang, “Interference-aware Topology Control

and QoS Routing in Multi-Channel Wireless Mesh Networks,” in Proc.of the 6th ACM international symposium on Mobile ad hoc networkingand computing, pp. 68-77, May. 2005.

[33] C.E. Perkins and E.M. Belding-Royer, “Quality of service for ad hocon-demand distance vector routing,” inInternet Draft, draft-perkins-manet-aodvqos-02.txt, October 2003, Work in progress.

[34] T.Liu and W. Liao “ ”Interference-aware QoS routing formulti-ratemulti-radio multi-channel IEEE 802.11 wireless mesh network,” in IEEETransactions on Wireless Communication, vol 8(1), pp.166-175, 2009.

[35] D. Ghosh, A. Gupta, and P. Mohapatra “ ”Admission controlandinterference-aware scheduling in multi-hop wimax networks,” in IEEEMASS, 2007.

[36] Y. Hu, X.-Y. Li, H.-M. Chen, and X.-H. Jia, “ ”Distributed call admissionprotocol for multi-channel multi-radio wireless networks,”in Proc. IEEEGlobecom, 2007

[37] M-Ali. Ergin, M. Gruteser, L. Luo, D. Raychaudhuri, H. Liu “ ”Availablebandwidth estimation and admission control for QoS routing inwirelessmesh networks,” inComputer Communications, pp.1301-1317, 2008.