department of computer science and engineering, charlotte ...science and engineering at johnson c....

Int. J. Ad Hoc and Ubiquitous Computing, Vol.

Multiple-metric hybrid anycast protocol for

heterogeneous access networks

Lijuan Cao*

Department of Computer Science and Engineering,Johnson C. Smith University,Charlotte, NC 28216, USAE-mail: [email protected]*Corresponding author

Kashif Sharif, Teresa A. Dahlberg and Yu Wang

Department of Computer Science,University of North Carolina at Charlotte,Charlotte, NC 28223, USAE-mail: [email protected]: [email protected] E-mail: [email protected]

Abstract: The wireless multihop to an Access Point (AP) model appears to be a promisingcomponent of future access network architectures. A key challenge is managing diverseresources at APs while discovering efficient multi-hop paths from source to AP basedon selection criteria dictated by applications or necessitated by network constraints.We propose a new hybrid proactive/reactive anycast routing protocol that integratesmultiple-metrics to calculate path cost and selects the appropriate AP. Simulation analysisshows that our approach outperforms single-metric protocols while supporting flexibleservice criteria, including load balancing at APs.

Keywords: anycast routing; multiple metrics; heterogeneous networks; hybrid routing.

Reference to this paper should be made as follows: Cao, L., Sharif, K., Dahlberg, T.A.and Wang, Y. (2011) ‘Multiple-metric hybrid anycast protocol for heterogeneous accessnetworks’, Int. J. Ad Hoc and Ubiquitous Computing, Vol.

Biographical notes: Lijuan Cao is an Assistant Professor in the Department of ComputerScience and Engineering at Johnson C. Smith University. Her research interest includeswireless sensor networks, peer-to-peer networks, heterogeneous access networks, resourcemanagement and routing algorithm design. She received PhD from University of NorthCarolina, Charlotte in 2008, and BS Degree from University of Electronic Science andTechnology in 2003.

Kashif Sharif is a PhD student in Department of Computer Science at the University ofNorth Caroline at Charlotte. His research focus is on routing and sink-discovery protocolsin application specific mobile ad hoc and sensor networks, path optimisations and resourcediscovery.

Teresa A. Dahlberg is a Professor of Computer Science at the University of North Carolinaat Charlotte. Her research on wireless networks addresses resource management protocols,data management for sensor networks, and analytic, simulation and experimental modelingand analysis techniques. She received the MS and PhD in Computer Engineering fromNorth Carolina State University and the BS in Electrical Engineering from the Universityof Pittsburgh.

Yu Wang received the PhD Degree in Computer Science from Illinois Institute ofTechnology in 2004, the BEng Degree and the MEng Degree in Computer Science fromTsinghua University, China, in 1998 and 2000, respectively. He is an Associate Professor ofComputer Science at the University of North Carolina at Charlotte. His research interestsinclude wireless networks, ad hoc and sensor networks, mobile computing, and algorithmdesign. He is a member of the ACM and a senior member of the IEEE, and IEEECommunications Society.

Copyright © 2011 Inderscience Enterprises Ltd.

37 Int. J. Ad Hoc and Ubiquitous Computing, Vol. 8, Nos. 1/2, 2011

1 Introduction

The vision of future generation networks is evolvingtowards one that includes interoperable heterogeneouswireless access technologies to provide seamless access tocore networks. Today’s markets include a proliferationof cellular, WiFi and WiMax technologies for access totelecom, internet and entertainment networks via mobiledevices such as phones, PDAs, laptops and sensors.Mobile devices equipped with multiple interfaces arealready available for consumer use. Recently, there havebeen active efforts to combine the advantages of cellularand ad hoc wireless access modes, and exploit their addedbenefits for system performance (Bhargava et al., 2004;Hsieh and Sivakumar, 2002; Liu et al., 2003; Ye et al.,2003; Wu et al., 2001; Zhou and Yang, 2002; Aggelouand Tafazolli, 2001; Wu et al., 2000; Lakkavalli et al.,2003; Yanikomeroglu, 2002; Kubisch et al., 2003). Muchof the proposed solutions comprise incremental changesto cellular resource management protocols or to ad hocrouting protocols as a way to extend cellular to ad hocor vice versa.

Here we take a more comprehensive approach byintegrating route discovery, access point discovery, andaccess point load balancing into a flexible multiple-metrichybrid routing protocol for heterogeneous wirelessaccess networks. Our design is based on our conclusionsthat future access networks will include heterogeneousair-interfaces, must support diverse applications, andwill have varying network resources and constraints.We therefore employ an anycasting routing paradigm,combined proactive and reactive route discovery,and multiple path cost metrics to support flexiblecost-performance decisions during route and access pointdiscovery.

To explain, first consider that within futuregeneration networks, mobile users may be able toaccess multiple Access Points (APs) for connection tothe internet. Thus, it is important for mobile users tolocate the best AP from “one or more of a group” ofAPs, which can be better modelled by an anycast ormanycast communications paradigm, rather than unicastor multicast. Here, the notion of ‘best’ can be describedas optimum for communication based on some selectioncriteria. As different applications might have differentrequirements, moreover, the cost of providing servicesto users varies from one AP to another, as determinedby a complex combination of issues including availablebandwidth, channel capacity, service availability, etc.The decision of AP selection should be determined byrouting method based on both application requirementand provided services from APs.

Though anycasting is originally an internet servicefor best effort delivery of datagrams from a host toat least one, and preferably only one server from thenearest ‘group of servers’, it has been applied to routingprotocol design for wireless ad hoc and sensor networks(Thepvilojanapong et al., 2005; Hou et al., 2005; Lenderset al., 2006; Intanagonwiwat and Lucia, 1999; Wang

et al., 2003; Peng et al., 2005). Wang et al. (2003)proposed anycast routing protocol based on Ad HocOn Demand Distance Vector (AODV) (Perkins et al.,2003). Anycast routing is supported by introducing a4-bit Anycast Group ID, which is contained in theRoute Request (RREQ) message along with other flagsfor discovery of the nearest anycast service provider.This protocol is designed to work purely in ad hocenvironments for evenly distributing the load on differentavailable anycast server nodes. In Peng et al. (2005),to support anycast service, Dynamic Source Routing(Johnson and Maltz, 1996) Protocol is extended with asimilar idea as Wang et al. (2003) for anycast ID orAnycast address.

Most existing routing protocols for multihop wirelessnetworks are simply designed using reactive schemes,meaning that route discovery is initiated on an as-neededbasis, for an initial connection, or when an existing routebreaks. In contrast, proactive routing protocols discoverand maintain routes before they are needed. Proactiveroute discovery minimises the time needed for pathselection, but requires additional overhead to maintainroutes that are not being used. Because of the need tominimise energy consumption and because of mobility,most ad hoc routing protocols are reactive. However,a unique feature of routing in heterogeneous wirelessnetworks, as compared to ad hoc routing, is that onlyAPs serve as possible ‘destinations’ to the multi-hop pathwithin the access network. These destinations are fixedand have specific access functionality. This calls for afresh look at the tradeoffs in using proactive or reactiveroute discovery policies.

Besides hybrid anycast routing paradigm, thecalculation of path cost is also a critical componentof routing design for heterogeneous access networks.In wireless networks, devices are usually resourceconstrained, e.g., limited battery capacity, buffer space,CPU processing capability, memory size, etc. Thus,the path cost metric, which guides path selection andresource consumption, is a crucial element of theprotocol design. Prior work on multi-hop wirelessrouting protocols relies largely on the use of single costmetric, such as number of hops in a path (Perkins andRoyer, 1999; Perkins and Bhagwat, 1994), the energyconsumed along a path (Singh and Raghavendra, 1998;Singh et al., 1998; Kim et al., 2002), the energy remainingafter using a path (Toh, 2001), the load carried by nodesalong a path (Lee and Gela, 2001; Wu and Harms,2001; Hassanein and Zhou, 2001; Saigal et al., 2004),etc. The aim of such protocols is to guide path selectionto favour the least-cost path, where the path cost metricreflects the criteria that is to be minimised. However,this approach does not suffice for future access networksfor at least two reasons: first, applications might havemultiple quality of service requirements that must besimultaneously considered during route discovery; andsecond, development of new radio access and wirelesscommunication technologies is producing a wide arrayof wireless devices, having different levels of constrained

Copyright © 2011 Inderscience Enterprises Ltd.

38 L. Cao et al.

resources. E.g., a laptop might have more powerfulprocess capability as compared to a cell phone.

In recent research, there has been some work focusingon optimal AP selection in IEEE 802.11 WLANs(IEEE Standard, 2008). In Vasudevan et al. (2005),Vasudevan et al. defined potential bandwidth as a metricto determine AP selection, which can be calculatedbased on delays experienced by beacon frames froman AP. In order to achieve load balancing amongmultiple available APs, Chen et al. (2006) presentedtwo algorithms to estimate the traffic load at APs byobserving the IEEE 802.11 frame delays and used theresult to determine the AP selection. In order to moreefficiently utilise the WLAN resource, Abusubaih et al.(2006) derived a metric based on traffic class generatedby the mobile node. In addition it takes into account thetraffic being generated by nodes in the same cell. Theauthors have shown through simulation that utilisingmultiple resource parameters, better throughput can beachieved. Rather than using certain specified metric tomake the AP selection decision, there is also somework using optimisation methodologies, for example,(Akl and Park, 2005; Lee et al., 2009) formulatedconstrained optimisation problems to achieve optimalAP selection and traffic allocation in the network. Whilethese work focus on a single hop environment, theprime difference of our work is the multihop nature.Moreover, rather than using a single metric to determinethe AP selection, our architecture has the ability toaggregate and transport multiple metrics for bothpath and AP selection. Nonetheless, these metrics andalgorithms can be adopted to work in our architectureeasily.

In this paper, we describe our hybrid anycastrouting protocol that integrates various cost metrics toguide path selection. Our hybrid mechanism divides themultihop portion of the access network into a proactiveand a reactive region. The proactive region enablesAPs to advertise their services by maintaining stateinformation at mobiles or relays within close proximityof the APs. The reactive region enables mobiles todiscover APs, as needed, by interrogating nodes inthe proactive region. A combination of proactive andreactive routing reduces communication overhead anddelay, while increasing throughput. The use of multiplecost metrics and anycast routing paradigm providesflexible support of service and resource requirements.

As our proposed solution is a general framework ofhybrid anycast protocol, it can be applied to a wide rangeof heterogeneous access networks (any wireless networkswith multihop relay to access points). These networksmay include combinations of cellular, Wimax, WiFi,mesh networks, sensor networks and other emergingpersonal and long distance communication standards,such as those systems in Bhargava et al. (2004), Hsiehand Sivakumar (2002), Liu et al. (2003), Ye et al. (2003),Wu et al. (2001), Zhou and Yang (2002), Aggelou andTafazolli (2001), Wu et al. (2000), Lakkavalli et al.(2003), Yanikomeroglu (2002) and Kubisch et al. (2003).

This paper is organised as follows. Section 2 describesthe system model and assumptions that we use foranalysis. Section 3 provides the details of our proposedmultiple-metric hybrid protocol while Section 4 presentssome issues in the implementation. Simulation results arepresented in Section 5. Section 6 concludes the paper. Apreliminary conference version of this paper appeared inCao et al. (2009). This version contains new analysis ofoptimal proactive radius, new experimental results, andbetter overall presentation.

2 System model

In this paper, we assume a general heterogeneousnetwork architecture as shown in Figure 1. There are twobasic entities in the system: Mobile Nodes (MNs) andAccess Points (APs). MNs are mobile devices which mayhave multiple interfaces (e.g., 3G, 802.11, or 802.16) aswell as the capability to relay traffic between interfaces.APs are physical access points that connect MNs to thecore network and terminate the wireless portion of thenetwork. Different APs can use various technologies. Weassume existing protocols or system designs (Bhargavaet al., 2004; Hsieh and Sivakumar, 2002; Liu et al.,2003; Ye et al., 2003; Wu et al., 2001; Zhou and Yang,2002; Aggelou and Tafazolli, 2001; Wu et al., 2000;Lakkavalli et al., 2003; Yanikomeroglu, 2002; Kubischet al., 2003) are available to integrate heterogeneousaccess technologies into the core network. We willonly focus on AP discovery and path selection in themultihop part of the architecture. We also assume thatall nodes have prior information about anycast groupmemberships and identities, using protocols available inthe literature.

Figure 1 Architecture of heterogeneous access network(see online version for colours)

3 Multiple-metric hybrid anycasting protocol

We derive our multiple-metric hybrid anycastingprotocol from the AODV architecture; with considerable

Multiple-metric hybrid anycast protocol 39

modifications to support anycasting, distributed regions,and multiple cost metrics. An objective of our protocolis for a mobile node to establish connection with anAP in an anycast group based on multiple path costmetrics, Thus, the selected AP can forward packets tothe destination in the core network.

3.1 Network regions

Hybrid routing has been studied in wireless adhoc networks. It leverages the tradeoff between thereduced delay provided by a proactive approach withthe reduced communications overhead provided by areactive approach. In order to support both proactiveand reactive approaches, the network is divided into tworegions:

• Proactive region. APs and MNs within an m hopradius of an AP are in the proactive region. AllMNs maintain active information about AP in thisregion through periodic Hello packets sent by AP.Hereafter, we call m the proactive radius.

• Reactive region. All MNs that more than m hopsaway from an AP are part of the reactive region,and use a reactive anycast routing protocol todiscover routes to an AP.

The objective of our hybrid anycast protocol is for amobile node to establish communication with an AP inan anycast group so that the selected AP can forwardpackets to the destination in the core network. The routeto destination for all data packets is selected throughany of the AP based on a decision metric. APs are entrypoints for MNs to access the internet and are part of oneor more anycast group(s).

3.2 Protocol functionality

Protocol functionality of our proposed anycast protocolcan be divided into the following phases.

• Hello message transmission. All APs periodicallytransmit HELLO, which only traverse m hops (i.e.,inside the proactive region), as defined by using theTTL value in the IP header. Upon receiving a Hellopacket, the route to the AP is created or updated,including the current capacity of the AP, as well asthe generic cost of the path to the AP. Only nodeswithin m − 1 hops distance from the AP decreasethe TTL value and rebroadcast the packet.

• Route discovery (Proactive region). An MNdetermines that it is in the proactive region if it hasreceived HELLO from any AP that belongs to thedestination anycast group in the previous RouteExpiration time interval. If the capacity of that APcan satisfy its application requirement, it can startsending data using the information in the routingtable without performing route discovery;otherwise, it performs route discovery as reactiveregion nodes.

• Route discovery (Reactive region). If an MN doesnot have any valid route available to any memberof the anycast group in its routing table, itbroadcasts a RREQ. Upon receiving the RREQ, anintermediate node first checks whether it hasreceived this RREQ before. If yes, it drops theRREQ. Otherwise, it updates the hop_count entryby adding one, and updates the cost field withequation (8). The intermediate node then creates anew entry in its routing table to record the previoushop and rebroadcasts the RREQ.RREP can only be generated by AP members ofthe anycast group or MNs in proactive regions thathave an active path to any member of the anycastgroup. Therefore, upon they receive the first RREQmessage, they check whether the AP’s capacity cansatisfy the requirement in the RREQ message. Ifthe requirement is satisfied, RREP is generated andsent back to the source along the reverse path;otherwise, RREQ message will be dropped. Forlater RREQ messages, RREP is only generated forthose with smaller path cost value. Upon receivingthe first RREP, an intermediate node records theprevious hop and relays the packet to the next hop.Similarly, later RREP message will be forwardedonly if it has a smaller path cost value.

• Route selection. Route selection is related to thecost metric used in the protocol, i.e., AODV selectsthe path with the first RREP. While usingmultiple-metric included in the RREQ, our anycastprotocol selects the route with the smallest costvalue out of all received RREPs. After the sourcenode receives the first RREP, it starts sending outdata. It will switch to the other path only if the costvalue in the corresponding RREP issmaller.

• Route maintenance. Route maintenance is the sameas for classical AODV.

3.3 Analysis of optimal proactive radius

Notice that the hybrid proactive/reactive approach inour hybrid anycast protocol can reduce overhead of APdiscovery. However, the radius m of proactive region isan important parameter, which can greatly influence thenetwork performance. Therefore, in this subsection, wefocus on analysis of the optimal m value in terms ofoverhead.

Before giving our theoretical analysis, we list theassumptions of our models:

1 APs and MNs have the same transmission range r

2 All APs belong to the same anycast group and theircoverage does not overlap with each other

3 MNs only rebroadcast a RREQ message once

4 Both MNs and traffic sources are uniformlydistributed in the network

40 L. Cao et al.

5 The network is distributed in a disk area withradius R.

Table 1 presents all the parameters we use in our analysis.

Table 1 Parameters in our analysis and their values in theplots

Symbols Meaning Values

R Network radius 2000mr Transmission range of an AP or MN 250mm Radius of proactive region of an AP Variouss Number of traffic sources in the

networkVarious

l Number of APs in the anycast group Variousd MNs distribution density 4 per m2

t Total time of operation 500 secondsη HELLO broadcast interval 20 seconds

The total overhead of AP discovery can be dividedinto two parts: Hello messages from APs inside theproactive region and RREQ messages from MNs insidethe reactive region. Here, we ignore the RREP messages,since they are sent along unicast routes which leads toa much lower number as compared to the number ofHELLO and RREQ messages which are sent by flooding.

Each HELLO message floods the proactive regionand it can reach m hops with m − 1 re-broadcasts.Therefore, the total number of HELLO messagesbroadcast per AP is

π((m − 1)r)2d, (1)

where d is the node density and π((m − 1)r)2 is thearea of the proactive region of this AP. Then, the totalnumber of HELLO messages from all l APs during thewhole operation is

l[π((m − 1)r)2d]t

η. (2)

On the other side, we calculate the total number ofRREQ messages. If the source node is located in theproactive region, there should be no RREQ overhead;otherwise, the number of RREQ messages per floodingfor one route discovery is

π(R2 − lm2r2)d. (3)

Here π(R2 − lm2r2) is the area of the reactive area.Notice that when a RREQ reaches the proactive regionof any AP in the anycast group, it will not be rebroadcastanymore. Since we assume s sources are uniformlydistributed in the network, the number of traffic sourceslocated in the reactive area is

sR2 − lm2r2

R2 . (4)

Thus, the total number of RREQ messages of all the ssources can be calculated by multiplying equations (3)and (4):

sR2 − lm2r2

R2 × π(R2 − lm2r2)d = πsd(R2 − lm2r2)2

R2 .

(5)

Therefore, the total number of overhead can be expressedas a function of m, (if HELLO and RREQ packets havethe same size)

f(m) = l[π((m − 1)r)2d]t

η+ πsd

(R2 − lm2r2)2

R2 . (6)

If HELLO and RREQ are not of the same size, we canmodify the above equation to:

f(m) = αl[π((m − 1)r)2d]t

η+ βπsd

(R2 − lm2r2)2

R2 ,

(7)

where α and β are the HELLO and RREQ packet size,respectively.

Using a common setting of the parameters, as shownin Table 1, we plot the overhead function of f(m) withdifferent combinations of l and s. Figure 2(a) shows theplot of f(m) with various numbers of sources when onlya single AP is inside the anycast group. The followingobservations are summarised:

• total overhead increases with the number of trafficsource increases, since the number of RREQmessages increases

• all curves follow the same trend, first decreasing tothe lowest point, then increasing, and mergingtogether when r = 8, where it is completelyproactive

• the optimal value of m (where f(m) is minimised)increases as the source number increases, i.e., 3 for10 sources, 5 for 15 sources, 6 for 20 and 25sources; this has been confirmed later in oursimulation.

Figure 2(b) shows a similar scenario with 4 APs. Thetrend of all curves is the same as that with single AP .Notice that the merge point of all curves shifts to 4, sincem = 4 can guarantee the proactive region covers most ofthe network. To observe how the quantity of AP affectsthe overhead function, we fix s = 20 and plot the set ofcurves with different l values in Figure 2(c). The curvesfollow the similar trend (first decreasing then increasing)as shown in Figure 2(a) and (b). It is interesting thatthe optimal value of m decreases as the number of APincreases. In other words, as more APs belong to thesame anycast group, each AP can reduce its proactiveradius.

3.4 Simulation study on proactive radius

We also conduct simulations with ns-2 (http://www.isi.edu/nsnam/ns) to test the performance of hybridanycasting with different proactive radii. Our simulationsuse the IEEE 802.11 Distributed Coordination Function(DCF) MAC protocol. 100 nodes are randomlydistributed in a 2200m × 1000m rectangular region.While 4 APs are fixed around the four corners, MNs


Figure 2 Overhead vs. proactive radius: plot of f(m):(a) 1 access point; (b) 4 access points and(c) 20 sources (see online version for colours)

move freely with a maximum speed of 20m/s usingRandom Way Point (RWP) mobility model (Johnsonand Maltz, 1996). Each round of simulation runs for500 seconds. We generate various mobility degreeswith different pause time values (0, 100, 200, 300, 400

seconds), varying from high mobility (low pause time) tovery low mobility (high pause time and almost static).Constant Bit Rate (CBR) sources are used, and thecommunication pairs are randomly chosen over thenetwork. 10, 20 and 30 sources are used to representdifferent load degrees, each sending 4 packets per secondwith size 512 bytes.

Figure 3 shows the simulation result, comparingthe performance of our hybrid protocol with differentproactive radii (0, 1, 2, 3 and 4). Due to space limitation,here we only present the delivery ratio and overheadmeasurements for four types of mobility. It is clear thataverage delivery ratio increases as pause time increases,also as the number of traffic sources decreases. InFigure 3(d), the delivery ratio of 10 traffic sources arelower than that of 20. This is because the network isnearly static when the pause time is 400 seconds, andthere might be a partition which causes lower deliveryratio in the network. Although there are fluctuationsin the plots, the main trend is coherent, increasingto a peak point, then decreasing, which means thatthe proper proactive radius selection can improve theperformance, i.e., in Figure 3(b), the optimal proactiveradius is 2. Figures in the lower row of Figure 3 comparethe normalised routing overhead of the same scenarios,which also increases as the pause time increases. Thisset of curves also show a rough trend, decreasing to thelowest point, then increasing, similar to the observationin the theoretical analysis in Section 3.3. For example,in Figure 3(d), with 10 traffic sources, overhead achievesthe lowest point when the proactive radius is 1, 20sources with 3, and 30 sources with 4. This also confirmsone of our conclusions from the theoretical analysis:optimal proactive radius increases as the number oftraffic sources increase. It is not possible to determinea particular radius value, as it varies for differentsituations. One of our future work is to devise algorithmsfor optimal radius determination at run time based onnetwork parameters.

3.5 Application requirements

Classical unicast routing protocols aim to discover theminimum cost path to the single destination specified bythe source node at network layer, for instance, AODVuses hop counts as the route metric. However, forheterogeneous access networks, anycast routingparadigm are required to locate the best AP among agroup as well as discover the minimum cost path to thatAP which can fulfill the various applicationrequirements. Here, we classify the applicationrequirements into two categories:

• Requirements for AP selection. Differentapplications might have different requirements forcertain types of resource at APs. On the other side,due to device or technology diversity, different APsmight have different levels of capability. Takingcapacity for example, APs utilising 3G access

42 L. Cao et al.

Figure 3 Average delivery ratio (the left column) and normalised overhead (the right column) with different proactive radii,number of sources, and pause time values: (a) pause time − 0 sec; (b) pause time − 200 secs; (c) pause time − 300 secs;(d) pause time − 400 secs; (e) pause time − 0 sec; (f) pause time − 200 secs; (g) pause time − 300 secs and (h) pausetime − 400 secs (see online version for colours)

technology can support simultaneous transmissionsto several users by assigning different channels tothem. Therefore, the system capacity, can bemodelled as maximum number of connections. Onthe other side, AP with 802.11 access technology

only provides one channel to all users by applyingDistributed Coordination Function (DCF) MACprotocol, which uses a contention algorithm toprovide access to all traffic. Thus, the current trafficload can be used as an indication for the capacity


of the AP, the higher the traffic load, the lessavailable resource, vice versa. To guarantee thequality of communication, AP selection mustsatisfy certain application requirements, i.e.,selected AP has enough resource for theapplication. Notice that there are various ways todefine or measure the resource at APs. Forsimplicity, we only use the number of connectionsat APs as an example of the metric for APselection. However, other metrics (such as the APselection metrics defined in Vasudevan et al. (2005),Chen et al. (2006) and Abusubaih et al. (2006)) canalso be used in our multi-metric anycast protocols.

• Requirements for path selection. As the multihoprelay service is provided by other MNs in thenetwork, which are usually resource constrained,for example, constrained battery capacity, limitedbuffer space, CPU processing capability, memorysize, etc. As availability of such resources cangreatly affect the performance of the connection,applications may require the route discovery toguarantee the sufficiency of certain resource on theselected route. Besides the availability of resources,applications may also require to minimise certaincost metric to optimise the performance, e.g., hopcount. On the other side, the ‘relay’ service is notfree as it causes resource consumption at mobileusers providing the service. Therefore, in order tocommunicate efficiently, applications might alsolimit certain cost metrics related to the expenditure,e.g., energy consumption. Thus, applicationsrequire using of multiple metrics for path costcalculation to guarantee the performance. Here,various route metrics at different layers (such asphysical layer or link layer) can be used. Theremight be multiple APs that can satisfy therequirements for AP selection, and even with onlyone AP, there might be multiple paths availablebetween source and the AP(s). Based on themultiple-metric path cost specified by theapplication requirement, path with the minimumcost value will be selected as the best route.

3.6 Multiple metrics path cost

Since applications require simultaneously using multiplemetrics to determine the path cost, we use a simple linearcombination of different routing metrics, as shown infollowing equation:

cost = cost′ +∑

∀i

αi × metrici (8)

where cost′ is the accumulated cost of previous nodesalong the path; metrici is scaled value from (0, 1);and αi is the weighted factors (or called coefficients)for metrici to calculate the cost. Based on applicationrequirement, these weighted factors can be flexibly variedto change the importance of the cost metrics during routediscovery.

In Alkahtani et al. (2006), the authors proposed toapply Analytic Hierarchy Process (AHP) (Saaty, 1980)for the calculation of combined four QoS metrics. Eventhough AHP can normalise the value of metrics from(0, 1) based on relevant costs among paths, it requiresroute discovery message to carry the multiple metrics.This causes more control overhead as well as space costat nodes to maintain the path information. Moreover,the complexity of the calculation is higher. Thus, in ourdesign we uses the linear combination. However, withsimple modification (additional fields in RREQ and AHPimplementation at APs and its proactive regions), AHPcan also be applied to our anycasting method.

Furthermore, in heterogeneous networks, differentmobile devices might have different levels of constraintfor the resources, for example, laptop might have morepowerful process capability compared to a cell phone.In order to fit the device diversity for heterogeneousnetworks, the protocol can also adaptively change theweight value based on the node class during the routediscovery phase. Device classes could, for example, bedifferentiated based on the battery type, the amount ofmemory, or the air interfaces present in a particular typeof mobile device.

4 Protocol implementation

In this section, we briefly introduce some detailed designsin the implementation of our proposed protocol.

4.1 Packet format

Four types of control packets are designed for theprotocol, as explained in this section.

• Hello packet(HELLO). This packet is a specialtype of packet generated only by the APs, which isbroadcast periodically inside the proactive region.As shown in Figure 4, the packet includes followingfields:

• Type: 1-byte field keeps the packet type.

• Life time: 1-byte field keeps the time (inseconds) during which this information is valid.

• Originator Anycast Group ID: 2-byte fieldwhich represents the AP’s anycast group.

• Originator IP Address: 4-byte unicast addressof the AP.

• Originator Sequence Number: 4-byte sequencenumber of the AP.

• Path Cost Metric: 4-byte value which caninclude a set of path cost values along the pathbetween the AP and the last hop node.

• Current Capacity: 4-byte value keeps thecurrent available capacity of the AP.

44 L. Cao et al.

• Route Request Packet (RREQ). For MNs that donot have any valid route available to any memberof the anycast group in its routing table, RREQpacket is generated to initialise the route discovery.Figure 5 shows the format of the packet, which issimilar to that of AODV protocol. The majordifferences are: instead of using unicast address asdestination address, the packet has the anycastgroup ID as the destination address; two morefields are added for adapting applicationrequirements and utilising multiple metrics as pathcost. Accumulative path cost is the generic pathcost to current node using equation (1). Genericapplication requirements includes bothrequirements for AP selection (e.g., capacity) andrequirements for path selection (e.g., the weightedfactors of each routing metrics).

• Route Reply Packet (RREP). This packet isgenerated by APs or MNs in proactive region forcorresponding RREQ packets. The format of thepacket, as shown in Figure 6, adds two more fieldscompared to that of AODV. While destinationanycast group ID represents the anycast group thatthe destination node belongs to, the accumulativepath cost is the accumulative cost along the pathfrom the destination node to the source node.

• Route ERROR Packet (RERR). The route errorpacket (RERR) is the same as that of AODVprotocol.

Figure 4 Hello packet

Figure 5 Route request packet

4.2 Cost metrics

We assume that different APs may have differentcapacity, i.e., they can only serve up to certain amount oftraffic. We use the total possible data rate at each AP asthe measurement of its capacity. In our simulation, for

Figure 6 Route reply packet

a simple demonstration, our protocol adopt three pathcost metrics: hop count, energy cost, and traffic load.

Traffic load: We use the interval time between receivingtwo data packets to estimate the traffic load of anode. To do that, each node maintains a weightedinterval value intval which is scaled and limited in arange of [0, 1], where 1 indicates no traffic load and 0indicates high load. Upon receiving a data packet, intvlis updated by intvl = (1 − β) × intvlold + β × intvlnew.Here intvlold/intvlnew are the old/new interval valuesand β is an adjustable parameter (in our simulation, β =0.2). Then the traffic load is 1 − intvl.

Energy cost: We consider the transmission power ateach node as the power cost metric. A simple powerconsumption model is used where the power consumedby a one-hop link uv is proportional to c||uv||4. Here||uv|| is the distance between nodes u and v, and c isa constant. Notice that remaining power capacity ormore complex power consumption model can also beused as a power cost metric, but here we only use thesimple transmission power to demonstrate the efficiencyof our combination metrics for energy aspect. In NS-2, the power consumed by the maximum transmissionrange 250 m is 0.66 watt per second, so the transmissionpower consumed by a link uv is 0.66 × ||uv||4/(250)4.Therefore, the range of power consumed each hop isfrom 0 to 0.66, and we can easily normalise this metricby dividing it by 0.66.

Therefore, the path cost equation becomes:

cost = cost′ + α1 × 1 + α2 × load + α3 × energy_cost.(9)

Here, ‘1’ is the hop count; load represents the trafficload at the current node; and energy_cost denotes thenormalised energy cost for the link from the previoushop to the current node. Different applications candefine their requirement by including different sets ofweighted values in RREQ. For example, an applicationmight only want to consider energy consumption, thus,(α1, α2, α3) = (0, 0, 1).

In our protocol, the path cost metric field in HELLOcontains the three cost metrics; while the accumulativepath cost field in RREQ contains the value of combinedpath cost from equation (9). In RREQ, the genericapplication requirement field includes two portions: oneis the required capacity (data rate), and the other hasthree weighted factors for path cost calculation.


5 Performance evaluation

In this section, we conduct several sets of simulationswith NS-2 (http://www.isi.edu/nsnam/ns) to evaluateour proposed multiple-metric hybrid protocol.Our simulations use the IEEE 802.11 DistributedCoordination Function (DCF) MAC protocol.

In order to demonstrate how different requirementsand path cost metrics guide route discovery and resourceconsumption, we conduct simulations with three differentnetwork deployment. Following metrics are used toevaluate the performance:

• Packet delivery ratio: ratio of the data packetsreceived at destinations to that sent out fromsources

• Average end-to-end delay: average time betweendata packets sent out from sources and received atdestinations

• Normalised routing overhead: ratio of controlpackets transmitted to data packets received atdestination

• Energy: total energy consumption at the end of thesimulation

• Throughput: the number of data packets received ateach destination per second.

5.1 Scenario 1: AP selection and different costmetrics

We first use a simple network with two APs todemonstrate the selection of APs based on AP capacity.The network deployment is shown in Figure 7, whereblack triangles are APs, blue dots represent relayingMNs, red dots represent source MNs who generatethe traffics, and circles around APs are their proactiveregions. Two APs AP0 and AP1 belong to a singleanycast group. There are two CBR traffic sources in thecentre of the network, both of which can connect to thetwo APs through multihop routes. It is easy to see thatPath 1 is the shorter path with only 4 hops and Path 2has less energy consumption. The data rate is 8 packetsper second for both sources. The two connections start at10 and 15 second, respectively. We set proactive regionradius as 2, therefore, both of the source nodes arelocated in reactive region. Here, we conduct two setsof simulations to demonstrate the route discovery withdifferent application requirements.

Figure 7 Network deployment for Scenario 1 (see onlineversion for colours)

In the first set of simulations, the capacity of APs areboth 12 packets per second, thus, each AP can onlyhandle one connection. We conducted two simulationswith and without requirement for AP selection, whilehop count is used as the cost metric for path selectionin both simulations. The simulation results show that: ifthe application specified capacity requirement, i.e., onlyAPs with enough capacity can correspond RREP, thefirst connection chooses AP0 (since Path 1 is shorterthan Path 2), then the second connection chooses AP1since AP0 does not have enough capacity; the packetdelivery ratio is 100%. Otherwise, if the protocol does notinclude capacity requirement as selection criteria, bothconnections choose AP0; the packet delivery ratio is only77.67%. The throughput of the two APs are shown inFigure 8. As we can see, the low delivery ratio is causedby traffic overflow at AP1.

Figure 8 Throughput of our protocol with (top) and without(bottom) requirements for AP capacity in the firstset of simulations in Scenario 1 (see online versionfor colours)

After demonstrating how requirements for AP selectionaffect the route discovery, let us study how differentpath cost metrics affect the simulation results.In the second set of simulation, the capacity areincreased to 20 packets per second for both APs.Therefore, both of the APs have enough capacity

46 L. Cao et al.

to process two connections simultaneously. Here,we ran three simulations with different requirementsfor path selection. As explained in Subsection 4.2,different requirements are specified by different sets

Figure 9 Throughput of our protocol with different pathcost metrics in the second set of simulationsin Scenario 1: (a) Hop count; (b) Traffic load and(c) TX power (see online version for colours)

of weighted values. In this set of simulation, we use(α1, α2, α3) = (1, 0, 0), (0, 1, 0), (0, 0, 1), respectively.Thus, the protocols use hop count, transmission power,and traffic load as application requirement, respectively.We compared the delivery ratio (DR), average end toend delay (Delay), normalised overhead (Overhead), andtotal consumed energy (Energy), as shown in Table 2.The throughput of the two APs are shown in Figure 9.

As we can see, the difference of delivery ratio andoverhead among the three simulations are not significant.In the simulation which used hop count as requirement,both of the connections choose path 1 (as shownin Figure 9(a)), which is shorter than Path 0. Thus,this simulation achieved the lowest delay compared toother two simulations, 10–20% lower. On the otherside, since Path 2 has a smaller energy consumption,in simulation using transmission power as requirement,both sources choose Path 2 to forward traffic (as shownin Figure 9(c)). Thus, this simulation has the minimumenergy consumption, 11–18% less. Also, as a result of thelonger path, it has the highest delay. The performanceof the simulation using load as requirements is betweenthe above two, since the first connection chooses Path 1,and the second connection chooses Path 2 as shown inFigure 9(b).

Table 2 Performance comparison of protocols utilisingdifferent path cost metrics in Scenario 1

Hop count Load Tx power

DR 100% 100% 99.83%Delay 0.028 0.030 0.035Overhead 0.211 0.203 0.208Energy 7.329 6.708 5.978

5.2 Scenario 2: Combining multiple metrics

In the second scenario, we consider combining multiplemetrics to find a path to an AP. As shown in Figure 10,the network includes one AP, five CBR traffic sources,as well as multiple paths from the sources to the AP.The sources start generating traffic one by one, andthe beginning point is (10, 15, 20, 25, 30) seconds,respectively. The data rate of sources is 4 packets persecond, and the capacity of the AP is 25 packets persecond which is sufficient for serving all traffics.

Figure 10 Network deployment of Scenario 2 (see onlineversion for colours)


We conducted four simulations with following setsof weighted values: (1, 0, 0), (0, 1, 0), (0, 0, 1) and(1/3, 1/3, 1/3) which represent hop count, load,transmission power, and combined metric of threemetrics. In order to study the route selection with thesedifferent requirements, we recorded the number ofpackets traversed through every intermediate node inreactive region. The result is shown in Figure 11. Clearly,when using hop count, the protocol mainly chooses MN10 to get connected to the proactive region, which makesthe route shortest. When using transmission power, thepackets are mainly forwarded through MNs 3 and 4,which can guarantee the minimum energy consumption.When using accumulative load as path selection criteria,the protocol tends to distribute traffic among multiplepaths. However, as a result of the accumulative manner,it favours path with smaller hop count if two paths carrythe same amount of load. Therefore, the traffic throughMN 10 is still high. On the other side, if multiple metricsare used as path selection criteria, the protocol can moreevenly distribute the traffic to some extent. Based onabove analysis of path selection, it is easy to understandthe result of the performance metrics shown in Table 3.

5.3 Scenario 3: More general case

Now we consider a more general complex network asshown in Figure 12, which includes two APs locatedon the left and right side. There are 5 CBR trafficsources that start generating traffic one by one, and thebeginning point is 10, 15, 20, 25, 30 seconds, respectively.

Figure 11 Traffic distribution among paths in reactive regionin Scenario 2 (see online version for colours)

Table 3 Performance comparison of protocols utilisingdifferent path cost metrics in Scenario 2

Hop count Load Tx power M-Metric

DR 99.68% 99.69% 99.38% 99.67%Delay 0.030 0.032 0.038 0.034Overhead 0.298 0.301 0.37 0.328Energy 3.864 3.716 2.673 3.245

The distance between sources and AP0 is smaller thanthat between sources and AP1. We vary APs’ capacityand the data rate of traffic sources to conduct threesets of simulations with this scenario. The configurationof these parameters are listed in Table 4. Within eachsimulation set, there are eight simulations that usevarious requirements to guide the route discovery, asshown in Table 5. For example, in Simulation 1, theapplication informs routing agent its data rate (as shownin the first row of Table 4) as well as the weightedvalues, which is (1, 0, 0); the second simulation used thesame weighted values, however, it does not include therequirement for AP selection.

The simulation results of Set 1 are shown in Table 6,and the throughput of each AP is shown in Figure 13.Obviously, the capacity of both APs is enough to hold

Figure 12 Simple network deployment in Scenario 3(see online version for colours)

Table 4 AP capacity and data rate of traffic sourcein Scenario 3

AP0 AP1 Src1 Src2 Src3 Src4 Src5

Set 1 25 25 4 4 4 4 4Set 2 15 15 4 4 4 4 4Set 3 10 50 4 5 8 4 5

Table 5 Requirements for AP and path selectionin Scenario 3

Requirements for Requirements forAP selection path selection

Simulation 1 Data rate (1, 0, 0)Simulation 2 NULL (1, 0, 0)Simulation 3 Data rate (0, 1, 0)Simulation 4 NULL (0, 1, 0)Simulation 5 Data rate (0, 0, 1)Simulation 6 NULL (0, 0, 1)Simulation 7 Data rate (1/3, 1/3, 1/3)Simulation 8 NULL (1/3, 1/3, 1/3)

Table 6 Performance comparison of simulations in Set 1of Scenario 3

DR (%) Delay Overhead Energy

Simulation 1 99.84 0.034 0.399 7.05Simulation 2 99.84 0.034 0.399 7.05Simulation 3 99.52 0.035 0.372 6.555Simulation 4 99.52 0.035 0.372 6.555Simulation 5 99.52 0.040 0.442 5.617Simulation 6 99.52 0.040 0.442 5.617Simulation 7 99.84 0.038 0.393 5.832Simulation 8 99.84 0.038 0.393 5.832

48 L. Cao et al.

all the traffic load, therefore, the requirement for APselection does not take effect, i.e., the results are thesame for the protocols that use the same weighted values,e.g., Simulation 1 and Simulation 2. As we can see fromFigure 13(a), nearly all the traffic in the first simulation

are directed to AP0 (except several single packets at thebeginning of the connections), which has the smaller costin hop count to the sources. We can observe the similartrend in Figure 13(e), because the energy cost is lessto connect to AP0 as compared to that of AP1. This

Figure 13 Throughput of simulations in Set 1 of Scenario 3: (a) Simulation 1;(b) Simulation 2; (c) Simulation 3; (d) Simulation 4;(e) Simulation 5; (f) Simulation 6; (g) Simulation 7 and (h) Simulation 8 (see online version for colours)


can also be verified from the performance metrics inTable 6: Simulation 1 achieves the smallest delay whileSimulation 5 obtains the minimum energy consumption.In Simulation 3, as shown in Figure 13(c), AP0 gotconnected with three traffic sources while AP1 got two.As we have explained in Scenario 2, accumulative loadmetric prefers shorter paths. The throughput of the APsin Simulation 7 is similar, three sources chose AP0 whilethe other three chose AP1.

The simulation result of the second set is shownin Table 7, and the throughput of the APs is shownin Figure 14. In this set of simulation, we have thesame data rate on the source side, but we decrease thecapacity to 15 for both APs. Therefore, at most, eachAP can only accommodate 3 connections simultaneously.In Simulation 1 and 5 (Figure 14(a) and (e)), weobserve that connections are distributed across twoAPs. Once AP0 reaches its maximum possible capacity,new connections are directed to AP1. In Simulation 2and 6 (Figure 14(b) and (f)) , as there is no capacityrequirements on APs, all the connections are formed withAP0 which has smaller cost. However, this reduces theoverall delivery ratio as shown in Table 7. Simulations 3,4, 7 and 8 (Figures 14(c), (d), (g) and (h)) depict similarbehaviour as in Set 1. Interestingly, the figures for loadmetric or combined metric with and without requirementfor AP selection are almost the same. This shows justusing load metric or combined metric can spread thetraffic among different APs.

In the third set of simulations, we considerheterogenous AP capacities and various traffic demands.The simulation result is shown in Table 8, and thethroughput of the APs is shown in Figure 15. Here, thedata rate for the five traffic sources are 4, 5, 8, 4 and5 packets, in the order of the start time of traffic. Thecapacity is configured as 10 and 30 packets for AP0 andAP1, respectively. Therefore, unlike previous simulationsettings (Sets 1 and 2), the connection distributionis different due to the different data rates and APcapacities. In Simulations 1 and 5 (Figure 15(a) and (e)),AP0 only accepts the first two connections. Simulation2 (Figure 15(b)) has the similar trend as compared tothat in Set 2, except that the upper bound is decreasedto 10 packets. In Simulation 6 (Figure 15(f)), due to theheavy traffic load along the path to AP0 which causesMAC layer collisions, the last connection is directed



Simulation 1 99.84 0.036 0.377 6.927Simulation 2 83.66 0.034 0.477 7.05Simulation 3 99.52 0.035 0.372 6.555Simulation 4 99.52 0.035 0.372 6.555Simulation 5 100 0.040 0.422 5.767Simulation 6 83.06 0.041 0.53 5.617Simulation 7 99.84 0.038 0.393 5.832Simulation 8 99.84 0.038 0.393 5.82

towards AP1. In Simulation 3 (Figure 15(c)), with thecapacity requirement, the connections are directed asfollow: sources 1 and 4 are connected to AP0 whilesources 2, 3 and 5 are connected to AP1. Simulation 7(Figure 15(g)) follows the same trend. Without limitationof capacity requirement, it is easy to understand thetrend in Figure 15(d) and (h) for Simulation 4 and 8.



Simulation 1 99.88 0.037 0.284 8.836Simulation 2 47.49 0.036 0.622 9.306Simulation 3 99.49 0.046 0.301 8.697Simulation 4 91.29 0.04 0.342 8.42Simulation 5 100 0.045 0.316 7.744Simulation 6 56.86 0.046 0.579 7.124Simulation 7 99.75 0.040 0.303 8.309Simulation 8 78.88 0.039 0.395 8.327

In summary, our simulations show different requirementsand routing metric can affect the AP selection androuting performances. Our multiple metric anycastrouting provide flexibility of picking appropriate metricsfor applications and users. On the other hand, APcapacity should be considered in the AP selection phaseto guarantee the performance.

6 Conclusion

We presented an anycast protocol for heterogeneousaccess networks to enable MNs to select one of multipleeligible access points. We further integrated a hybridproactive and reactive approach to AP discovery, whichsignificantly reduces the communication overhead. Thetheoretical analysis shows that the selection of theproactive radius can affect the network performanceand thus an optimal proactive radius is derived. Wealso conducted a set of simulations to evaluate theperformance of the protocol, and as an extension, weutilise traffic load as the cost metric for AP selection,and let APs dynamically adjust their attitudes fordissemination of HELLOs and acceptance of RREQsregarding their load states. The simulation results showthat our protocol effectively improves the performanceand provides the provision for load balancing and highservice availability.

To satisfy various application requirements (bothAP capacity and path metrics), we used combinedmultiple-metric to guide the route discovery and APselection. A set of simulations has been conducted toevaluate the performance of the proposed protocol.Simulation results showed that the utilisation ofmultiple-metric is important and effective to guide routediscovery and AP selection in heterogeneous wirelessaccess networks.

Further research work is in progress for building of adynamic algorithm to increase and decrease AP flooding

50 L. Cao et al.

Figure 14 Throughput of simulations in Set 2 of Scenario 3: (a) Simulation 1; (b) Simulation 2; (c) Simulation 3; (d) Simulation 4;(e) Simulation 5; (f) Simulation 6; (g) Simulation 7 and (h) Simulation 8 (see online version for colours)


Figure 15 Throughput of simulations in Set 3 of Scenario 3: (a) Simulation 1; (b) Simulation 2; (c) Simulation 3; (d) Simulation 4;(e) Simulation 5; (f) Simulation 6; (g) Simulation 7 and (h) Simulation 8 (see online version for colours)

52 L. Cao et al.

radius, based on network conditions. Also, furtheranalysis can be done on the effect of heterogeneoustransmission ranges of devices, as well as coverageoverlapping of APs.

Acknowledgment

The work of Yu Wang is supported in part by theUS National Science Foundation (NSF) under GrantNo. CNS-0721666, CNS-0915331, and CNS-1050398,and by Tsinghua National Laboratory for InformationScience and Technology (TNList).

References

Alkahtani, A.M.S., Woodward, M.E. and Al-Begain, K. (2006)‘Prioritized best effort routingwith four quality of servicemetrics applying theconcept of the analytic hierarchyprocess’, Computers & Operations Research, Vol. 33,3 March, pp.559–580.

Abusubaih, M., Gross, J., Wiethoelter, S. and Wolisz, A.(2006) ‘On access point selection in IEEE 802.11wirelesslocal area networks’, Proc. 6th InternationalWorkshop on Wireless Local Networks, November,Tampa, FL, pp.879–886.

Aggelou, G. and Tafazolli, R. (2001) ‘On the relayingcapabilityof next- generation GSM cellular networks’, IEEEPersonal Communications, Vol. 8, February, pp.40–47.

Akl, R. and Park, S. (2005) ‘Optimal access point selectionand traffic allocation in IEEE 802.11 networks’, Proc.9th World Multiconference on Systemics, Cybernetics andInformatics (WMSCI 2005): Communication and NetworkSystems, Technologies and Applications, Vol. 8, July,pp.75–79.

Bhargava, B., Wu, X., Lu, Y. and Wang, W. (2004) ‘Integratingheterogeneous wireless technologies: a cellular aidedmobile ad hoc networks’, ACM Mobile Network andApplications, Vol. 9, No. 4, August, pp.393–408.

Cao, L., Sharif, K., Wang, Y. and Dahlberg, T.(2009) ‘Multiple-metric hybrid routing protocol forheterogeneous wireless access networks’, Proc. 6thIEEEConsumer Communications and NetworkingConference, Las Vegas, NV, pp.1–5.

Chen, J., Chen, T., Zhang, T. and Berg, E. (2006) ‘WLC19-4:effective AP selection and load balancing in IEEE802.11wireless LANs’, IEEE Global TelecommunicationsConference, San Francisco, CA, pp.1–6.

Hassanein, H.S. and Zhou, A. (2001) ‘Routing withloadbalancing in wireless ad hoc networks’, Proc. 4thACM International Workshop on Modeling, Analysis andSimulation of Wireless and Mobile Systems, Rome, Italy,pp.89–96.

Hou, Y., Shi, Y. and Sherali, H. (2005) ‘On base stationselection for anycast flow routing in energy-constrainedwireless sensor networks’, Second InternationalConference on Quality of Service in HeterogeneousWired/Wireless Networks, Florida, August.

Hsieh, H-Y. and Sivakumar, R. (2002) ‘On using thead-hoc network model in cellular packet data networks’,MobiHoc ‘02: Proceedings of the 3rd ACM InternationalSymposium on Mobile Ad Hoc Networking & Computing,June, Lausanne, Switzerland, pp.36–47.

IEEE Standard (2008) ‘IEEE Standard: Wireless LANMedium Access Control (MAC) and Physical Layer(PHY) Specifications Amendment 1: Radio ResourceMeasurement of Wireless LANs’, IEEE Std 802.11k-2008 (Amendment to IEEE Std 802.11-2007), 12 June,pp.c1–222.

Intanagonwiwat, C. and Lucia, D. (1999) The Sink-BasedAnycast Routing Protocol for Ad Hoc Wireless SensorNetworks, Department of Computer Science, USC, Tech.Rep.

Johnson, D. and Maltz, D. (1996) ‘Dynamic source routing inad hoc wireless networks’, Mobile Computing, Vol. 353,pp.153–181.

Kim, D., Garcia-Luna-Aceves, J., Obraczka, K., Cano, J. andManzoni, P. (2002) ‘Power-aware routing based on theenergy drain rate for mobile ad hoc networks’, Proc. IEEEInternational Conference on Computer Communicationand Networks, October, Miami, FL, pp.565–569.

Kubisch, M., Mengesha, S., Hollos, D., Karl, H. andWolisz, A. (2003) ‘Applying ad hoc relaying to improvecapacity,energy efficiency, and immission in infrastructure-based WLANs’, Proc. Communications in DistributedSystems (KiVS), Germany, February, pp.195–206.

Lakkavalli, S., Negi, A. and Singh, S. (2003) ‘Stretchablearchitectures for next generation cellular networks’,International Symposium on Advanced RadioTechnologies(ISART), March.

Lee, J., Han, B., Lim, H., Kim, Y., Saxena, N. and Chung, T.(2009) ‘Optimizing access point allocation using geneticalgorithmic approach for smart home environments’, TheComputer Journal, Vol. 52, No. 8, Oxford UniversityPress, pp.938–949.

Lee, S.J. and Gela, M. (2001) ‘Dynamic load-aware routingin ad hoc networks’, Proc. of IEEE Int’l Conf. onCommunications (ICC), Vol. 10, June, Helsinki, Finland,pp.3206–3210.

Lenders, V., May, M. and Plattner, B. (2006) ‘Density-based vs. proximity-based anycast routing for mobilenetworks’, 25th International Conference on ComputerCommunications, Barcelona, Spain, April, pp.1–13.

Liu, B., Liu, Z. and Towsley, D. (2003) ‘On the capacityof hybrid wireless networks’, Proceedings of INFOCOM:22nd Annual Joint Conference of the IEEE Computer& Communication Societies, Vol. 2, San Francisco, CA,April, pp.1543–1552.

Peng, G., Yang, Y. and Gao, C. (2005) ‘ARDSR: an anycastrouting protocol for mobile ad hoc network’, Proceedingsof the IEEE 6th Circuits and Systems Symposiumon Emerging Technologies: Frontiers of Mobile andWireless Communication, Shanghai, China, Vol. 2, June,pp.505–508.

Perkins, C., Belding-Royer, E. and Das, S. (2003) Ad HocOn-Demand Distance Vector Routing, IETF RFC 3561.

Perkins, C.E. and Royer, E.M. (1999) ‘Ad-hoc on demanddistance vector routing’, Proceedings of IEEE WMCSA’99,New Orleans, LA.


Perkins, C. and Bhagwat, P. (1994) ‘Highly dynamicdestination-sequenced distance-vector routing (DSDV)for mobile computers’, ACM Conf. on CommunicationsArchitectures, Protocols & Applications, October,London, UK, pp.234–244.

Saigal, V., Nayak, A.K., Pradhan, S.K. and Mall, R. (2004)‘Load balanced routing in mobile ad hoc networks’,Computer Communications, Vol. 27, 3 February,pp.295–305.

Saaty, T.L. (1980) The Analytic Hierarchy Process: Planning,Priority, Setting and Resource Allocation, 1st ed.,McGraw-Hill, Columbus, OH.

Singh, S. and Raghavendra, C.S. (1998) ‘PAMAS-poweraware multi-access protocol with signaling for ad hocnetworks’, ACM SIGCOMM Computer CommunicationReview, Vol. 28, 3 July, pp.5–26.

Singh, S., Woo, M. and Raghavendra, C.S. (1998) ‘Power-aware routing in mobile ad hoc networks’, Proc.MobiCom, October, Dallas, TX, pp.181–190.

Thepvilojanapong, N., Tobe, Y. and Sezaki, K. (2005) ‘HAR:Hierarchy-based anycast routing protocol for wirelesssensor networks’, The 2005 Symposium on Applicationsand the Internet, Italy, February.

Toh, C.K. (2001) ‘Maximum battery life routing tosupport ubiquitous mobile computing in wireless ad hocnetworks’, IEEE Communications Magazine, Vol. 39,No. 6, June, pp.138–147.

Vasudevan, S., Papagiannaki, K., Diot, C., Kurose, J. andTowsley, D. (2005) ‘Facilitating access point selection inIEEE 802.11 wireless networks’, IMC ‘05: Proceedingsof the 5th ACM SIGCOMM Conference on InternetMeasurement, October, Berkeley, CA, pp.293–298.

Wang, J., Zheng, Y. and Jia, W. (2003) ‘An AODV-based anycast protocol for mobile ad hoc network’,14th IEEE Proceedings on Personal, Indoor and MobileRadio Communications, Beijing, China, Vol. 1, September,pp.221–225.

Wu, K. and Harms, J. (2001) ‘Load-sensitive routing formobile ad hoc networks’, Proc. IEEE InternationalConferenceon Computer Communication and Networks,October, Scottsdale, AZ, pp.540–546.

Wu, X., Chan, S-H. and Mukherjee, B. (2000) ‘MADF: anovel approach to add an ad-hoc overlay on a fixedcellular infrastructure’, IEEE Wireless Communicationsand Networking Conference, Chicago, IL, Vol. 2,September, pp.549–554.

Wu, H., Qiao, C., De, S. and Tonguz, O. (2001) ‘Integratedcellular and ad hoc relaying systems: ICAR’, IEEEJournal on Selected Areas in Communications, Vol. 19,No. 10, October, pp.2105–2115.

Yanikomeroglu, H. (2002) ‘Fixed and mobile relayingtechnologies for cellular networks’, 2nd Workshop onApplications and Services in Wireless Networks, France,July, pp.75–81.

Ye, Z., Krishnamurthy, S. and Tripathi, S. (2003)‘A framework for reliable routing in mobile ad hocnetworks’, Proceeding of INFOCOM: 22nd Annual JointConference of the IEEE Computer & CommunicationSocieties, Vol. 1, April, pp.270–280.

Zhou, J. and Yang, Y. (2002) ‘Parcels: pervasive ad-hocrelaying for cellular systems’, Proceedings of MedHocNet,Sardegna, Italy, September.

Website

The network simulator – ns-2, http://www.isi.edu/nsnam/ns/

department of computer science and engineering, charlotte ...science and engineering at johnson c....

Documents