A swarm intelligence-based unicast routing protocol for hybrid ad hoc networks

ARTICLE IN PRESS

Journal of Systems Architecture xxx (2006) xxx–xxx

www.elsevier.com/locate/sysarc

ANSI: A swarm intelligence-based unicast routingprotocol for hybrid ad hoc networks q,qq

Sundaram Rajagopalan *, Chien-Chung Shen

DEGAS Networking Group, Department of Computer and Information Sciences, University of Delaware,

Newark, DE 19716, USA

Abstract

We present a hybrid routing protocol for both pure and hybrid ad hoc networks which uses the mechanisms of swarmintelligence to select next hops. Our protocol, Ad hoc Networking with Swarm Intelligence (ANSI), is a congestion-awarerouting protocol, which, owing to the self-organizing mechanisms of swarm intelligence, is able to collect more informationabout the local network and make more effective routing decisions than traditional MANET protocols. Once routes arefound, ANSI maintains routes along a path from source to destination effectively by using swarm intelligence techniques,and is able to gauge the slow deterioration of a link and restore a path along newer links as and when necessary. ANSI isthus more responsive to topological fluctuations. ANSI is designed to work over hybrid ad hoc networks: ad hoc networkswhich consist of both lower-capability, mobile wireless devices and higher-capability, wireless devices which may or maynot be mobile. In addition, ANSI works with multiple interfaces and with both wired and wireless interfaces.

Our simulation study compared ANSI with AODV on both hybrid and pure ad hoc network scenarios using both TCPand UDP data flows. The results show that ANSI is able to achieve better results (in terms of packet delivery, number ofpackets sent, end-to-end delay, and jitter) as compared to AODV in most simulation scenarios. In addition, ANSI achievesthis performance with fewer route errors as compared to AODV. Lastly, ANSI is able to perform more consistently, con-sidering the lower variation (measured as the width of the confidence intervals) of the observed values in the results of theexperiments. We show that ANSI’s performance is aided by both its superior handling of routing information and also itscongestion awareness properties, though we see that congestion awareness in ANSI comes at a price.� 2006 Elsevier B.V. All rights reserved.

Keywords: Swarm intelligence; MANET; Hybrid network; Hybrid routing; Congestion aware routing

1383-7621/$ - see front matter � 2006 Elsevier B.V. All rights reserved

doi:10.1016/j.sysarc.2006.02.006

q A section of this work was presented at ICAI 2005, June 2005,Las Vegas, NV, USA.qq This work is supported in part by National Science Foun-dation under grant ANI-0240398.

* Corresponding author. Tel.: +1 302 831 1131; fax: +1 302 8318458.

E-mail address: [email protected] (S. Rajagopalan).

1. Introduction

Hybrid ad hoc networks consist of a mixture ofmobile, ad hoc network (MANET) nodes and nodeswhich belong to highly capable infrastructure suchas mesh networks or cellular networks. The problemof hybrid ad hoc networks is to make these networkswork efficiently without relying on pre-configured

.

mailto:[email protected]

2 S. Rajagopalan, C.-C. Shen / Journal of Systems Architecture xxx (2006) xxx–xxx

ARTICLE IN PRESS

network topologies or centralized control. Hybridad hoc networks are useful in many situations whereimpromptu communication facilities are requiredsuch as battlefield communications, and disasterrelief missions.

Since the problem of hybrid ad hoc networkingshares a lot of problems with typical MANET prob-lems, typical routing solutions for hybrid networksstart with a MANET routing solution and thenapply some optimizations to work for specific sce-narios. A number of ad hoc routing protocols havebeen proposed, for example [1–5], of which some ofthem, like AODV [1] work on hybrid ad hoc net-works. In proactive protocols such as [5], nodes inthe network maintain routing information to allother nodes in the network by periodically exchang-ing routing information. Nodes using reactive pro-tocols, such as [1,2], delay the route acquisitionuntil a demand for a route is made. Hybrid proto-cols, like [4,6], use a combination of both proactiveand reactive activities to gather routes to the desti-nations in a network—nodes using ZRP, for exam-ple, proactively collect routes in their zone, andother routes are collected reactively. In [6], on theother hand, the level of proactive activity and reac-tive activity are chosen autonomously by the nodesin the network, and proactive activity is only seenaround favorite destination nodes. In most tradi-tional reactive protocols, like [1,2], only when aroute breaks irreparably does the protocol mecha-nisms repair the damage. In reality, route deteriora-tion in mobile networks is most often not suddenbut gradual,1 and most often available routes getbetter/deteriorate gradually and not suddenly. Sothe routing protocol should continuously maintaininformation about the nodes in the local area to per-form effectively and avoid too may link breakages.

In this paper, we present a hybrid routing suite(with both proactive and reactive components) forhybrid ad hoc networks which uses the mechanismsof swarm intelligence [7] to select good routes to des-tinations. We use Swarm Intelligence (SI) because SImechanisms allow for self-organizing systems [8] andmaintain state information about the neighboringnetwork better than traditional MANET routingmechanisms. Self-organizing systems are robustenvironments where erroneous system behavior iscorrected autonomously by the coordinated working

1 Some routes, such as routes to neighbors, break suddenly,when the neighbors go out of range. We are commenting on thegeneral case here.

of lower-level components. The combination/inter-action of lower-level components in SI such aspositive/negative feedback and amplification of fluc-tuations along with multiple interactions are themechanisms which allow a node to change routinginformation quickly and efficiently to adjust to anever-changing local topology and route deteriora-tion, thus initiating fewer link breakages.

Our protocol, ANSI, uses a highly flexible costfunction which allows it to use the information col-lected from the local ant activity, such as the conges-tion status of the neighboring nodes, in useful ways.In addition, the ant-like working of our protocolallows for the maintenance of multiple routes to adestination. In nodes which use proactive routingin ANSI, this fact is used to perform stochastic rout-ing, and in nodes that use perform reactive routing(pure MANET nodes), when one route fails, othersmay be used. Our motivation comes from the factthat different networks face different conditions,and thus a protocol suite should allow for variousconfigurations as the network conditions dictate.Furthermore, supporting multiple routes simulta-neously is essential to ensure survivability of the net-work [9]. ANSI facilitates ad hoc unicast routing byexploiting route finding behaviors that are emergent

from ant packets working collectively, rather thanexplicitly coding them to cope with the problem.We formulate the routing problem at node i as aset of ‘‘food foraging’’ problems from nest i, whereeach ‘‘food source’’ is a destination d in the net-work. In this formulation, next hops are evaluatedon the basis of the strength of the pheromone trail2

on the link connecting a node and a next hop.The remainder of this paper is organized as fol-

lows: In the next section, we discuss a number ofapproaches and protocols which are related to ourresearch. In Section 3, we describe in detail the com-ponents of ANSI unicast routing protocol, and fol-low it with Section 4 where we discuss the results ofthe comparison of simulated models of ANSI witha popular routing protocol, AODV [1]. We concludein Section 5 with a brief note on future research effort.

2. Related work

The main ingredients of SI, positive/negativereinforcement, and amplification of fluctuations

2 The computational equivalent of the chemical deposited onthe forest floor by ants.

S. Rajagopalan, C.-C. Shen / Journal of Systems Architecture xxx (2006) xxx–xxx 3

ARTICLE IN PRESS

are achieved in an environment with multiple inter-actions among nodes [7]. Because the above compo-nents of SI are the lower-level components requiredfor self-organizing behavior, the benefits of usingSI-based algorithms are not fully accrued if theany of the above lower-level components are notpresent in a swarm intelligence-based protocol. Thisis because in any SI-based system, these aspectswork together in learning about the network.

In [10] Baras and Mehta describe a swarm intel-ligence-based reactive ad hoc routing protocolcalled PERA. PERA uses broadcast forward antsas exploratory agents sent out on-demand to findnew routes to destinations. Each ant holds a list ofnodes that were visited while exploring the network,and since these ants are broadcast at each node, aforward ant can result in several backward ants—ants sent by destination nodes in response to for-ward ants. This uncovers several routes for each for-ward ant sent, and at each node these multipleroutes found to the destinations are maintained asprobability values. As with AntNet [11], the routingtable Ri at node i is a probability matrix with aprobability entry Pijd as the probability that a datapacket at i’s FIFO queue will take the next hop j

to be routed to d. Positive reinforcement is managedin PERA using forward/backward ants and nega-tive reinforcement is implicit—no explicit aging ofthe pheromone trails is done. After a route has beenestablished, PERA regularly uses forward ants tofind newer routes to destinations. This is wasteful,considering the fact that forward ants cause a lotof network resources to be consumed and shouldnot be sent when not necessary.

In [12], Camara and Loureiro outline a sourcerouting scheme in which the network relies on loca-tion information and support from fixed infrastruc-ture. Owing to a source routing approach, thealgorithm relies heavily on a source M destinationroute which is available at the time of message cre-ation. New nodes in the network start with usingtheir neighbor’s routing table. The routing table,generated using shortest path algorithms, on theother hand, may contain information which is out-dated. Ants are unicast from a source to specific des-tinations, for example, the destination node may bethe node with the oldest information in the routingtable. This mechanism is used to make sure that therouting information in the source is updated andrecent. Thereby, ants are used in [12] with thesemantics of routing information updates, like clas-sical distance vector protocols such as DSDV or

DBF—ants are not used as feedback agents to rein-force routes positively (in the case when a route isstill good), negatively (when a route is no longergood) or explore new routes randomly—ants in thisapproach are unicast to specified direction, notallowing for amplification of fluctuations, anddepending on known metrics such as timestamp ofa route in the routing table.

The approach used in [13] by Heissenbttel andBraun also relies on location information, and is apurely proactive routing approach based on divid-ing the network into logical zones and assigninglogical routers to each. Ants—forward ants andbackward ants—are used by logical routers in thisapproach to periodically check if the logical linksconnecting it to a randomly chosen destination arefunctional and reflect on the current state of the net-work surrounding the logical router. Positive andnegative reinforcement are achieved by means ofmultiple interactions and pheromone additions (byforward and backward ants) and pheromone aging,respectively. Random amplification of a new goodroute in the face of topological fluctuations ispossible by random dissemination of ants to desti-nations.

In [14], Gunes et al. outline ARA, a multipath,purely reactive scheme. ARA uses forward antsand backward ants to create fresh routes from anode to a destination. When routes to a destinationD are not known at S, a forward ant is broadcast,taking care to avoid loops and duplicate ants. Whena forward ant is received at an intermediate node X

via node Y, the ant reinforces the link XY in X toroute to all the nodes covered so far by the forwardant. When a forward ant is received at D, a back-ward ant is created which backtracks the path ofthe corresponding forward ant. At each node thebackward ant is received, the link via which thebackward ant is received is reinforced, like the for-ward ant does, for all nodes which have been visitedby the backward ant. In ARA, data packets per-form the necessary (positive) reinforcement requiredto maintain routes. When a path is not taken, it sub-sequently evaporates (negative reinforcement) andcannot be taken by subsequent data packets. Underthe described scheme, amplification of topologicaland network fluctuations is not possible exceptunder extreme conditions when routes breakoften.

In [15,16], Di Caro et al. describe AntHocNet, ahybrid, stochastic approach to the routing problemin MANET. AntHocNet is a congestion-aware pro-


ARTICLE IN PRESS

tocol which only finds routes on-demand, but once aroute is established, the route is proactively main-tained. This approach, argued by the authors tobe more ant-like [16] than other competing ant-based protocols, will fail to reduce overheads in veryhigh traffic/mobility scenarios, owing to the rate atwhich proactive ants are potentially unicast whenthe mobility increases. This is because in high mobil-ity/traffic scenarios, routes get invalidated often andproactive activity has to increase appropriately tokeep a valid view of the network for routing, thusincreasing the load placed on the network. Indeed,we do agree with the comment that the authors ofAntHocNet make regarding the repeated path sam-pling, and ANSI manages to steer clear fromrepeated path sampling by carefully choosing whento engage in route discovery activity.

In [17], Wedde et al. present a new routingalgorithm for energy efficient routing in mobilead hoc networks. In their approach, they showthat BeeAdHoc, a reactive source-routing protocolinspired by the foraging principles of honey bees, isable to achieve energy consumption characteristicsas compared to DSR, AODV and DSDV withoutcompromising on traditional performance metricssuch as packet delivery and throughput.

Our protocol, ANSI, is a hybrid protocol pro-posed for hybrid ad hoc networks. Some character-istics seen in traditional on-demand routingprotocols can be seen in ANSI. For example, anoptimization used in AODV, expanding ring search,is also used in ANSI, albeit more efficiently, owingto the use of history information. Unlike traditionalMANET protocols which engage in route mainte-nance/discovery activity only when links break,ANSI continuously updates a node’s neighborhoodinformation using data packets and control packetsto alleviate the negative effects due to flooding thenetwork with route discovery/maintenance. In addi-tion, unlike traditional MANET protocols, ANSIhas a flexible cost function which allows it to per-form metric-centered routing. In our implementa-tion, we have performed congestion-aware routing,but it is easy to see how this cost function can bemodified to perform, say, energy efficient routing.

When compared to other ant algorithms forMANET routing, we note that to the best of ourknowledge, there exists no other ant algorithm forhybrid ad hoc networks, but ANSI is able to per-form well in both pure MANET and hybrid adhoc networks. In addition, the ANSI design under-stands the advantages of proactive/stochastic rout-

ing in immobile, highly capable infrastructure andapplies it only in those nodes, rather than lettingpure MANET nodes incur the costs due to the sameunder high mobility conditions. Lastly, the flexiblecost function (specifically, the congestion-awarenessproperty) in ANSI leverages the inherent nature ofswarm intelligence by collecting multiple routesand using them to perform load balancing in allsections of the network. This, as we will see, alsoalleviates the tendency to create hotspots in thenetwork.

3. ANSI unicast routing protocol

3.1. Protocol overview

ANSI is a hybrid routing protocol for hybrid ad

hoc networks comprising of both proactive and reac-tive routing components. Pure MANET (mobile)nodes in ANSI use only reactive routing, and chooseroutes deterministically, while nodes belonging tomore capable, infrastructured (immobile) networksuse a combination of both proactive and reactiverouting and perform stochastic routing when multi-ple paths are available. The outline of the processof ANSI routing is as follows:

1. When a route to a destination D is required, butnot known at a node S, S broadcasts a forward

reactive ant to discover a route to D.2. When D receives the forward reactive ant from S,

it source-routes a backward reactive ant to thesource S. The backward reactive ant updatesthe routing table of all the nodes in the path fromS to D, allowing for data transfer from S to D.

3. When a route fails at an intermediate node X, Xfirst checks if there are other routes which can beused to route the packet to D. If not, then ANSIbuffers the packets which could not be routedand initiates a route discovery to find D by usinga forward reactive ant to perform local routerepair. Additionally, X sends a route error mes-sage back to the source node S.

4. Nodes belonging to more capable, infrastructurednetworks maintain routes to their connected com-ponents proactively, by periodic routing updatesusing proactive ants. Nodes belonging to morecapable, infrastructured networks also use sto-chastic routing when multiple paths are available.In addition, each node in the infrastructurecollects information about which mobile nodesare connected to which infrastructure node.


ARTICLE IN PRESS

5. When a route at D is known at a MANET nodeS, ANSI deterministically chooses the best nexthop to reach the destination. If S is part of ahighly capable infrastructure, then S may chooseto perform stochastic routing to the destinationD, depending on the availability of multipathroutes.

We claim ANSI will perform better than typicalMANET protocols because of the working of theSI mechanisms at each node, which maintain rout-ing information and local information more effec-tively than traditional MANET routing protocols.In addition, the congestion-awareness of ANSI alsohelps in controlling the extent of congestion in hightraffic scenarios. Lastly, in hybrid networks, ANSIis able to leverage the power of nodes belongingto more capable networks to assist in routing activ-ities of the network. In Sections 3.3.1–3.3.4, weexplain the details of the above actions, and showhow the SI mechanisms work at each node in main-taining routing information.

3.2. Protocol model

3.2.1. Data structures

Data structure (1) below is the ant structure, car-ried by all ants, and data structures (2) and (3)below are maintained at each node, and are updatedevery time an ant arrives at the node.

(1) Ant structure: The following information is car-ried by an ant p:(a) The ant ID of the ant, which is the (node

ID, sequence number) pair.(b) The number of nodes, m, which p visits,

including the node p originated from.(c) The nodes-visited-stack (adapted from

[11]), Sp, containing information aboutnodes V = {v1,v2, . . . ,vm}, that can bereached by backtracking the ant p’s move-ment (using the nodes-visited-stack), and

(d) The pheromone amount at v 2 V, pv.
(2) Ant decision table at node i, Ai: (adapted from
[18]). An ant decision table is a data structurethat stores pheromone trail information forrouting from node i to a destination d via k pos-sible next hop nodes J = {j1, j2, . . . , jk}. The linkij in the ANSI network, between two nodes i

and j is assumed to be bidirectional. Routingtables are computed from ant decision tables.Each ant decision table entry Aijd for node i

maintains a row for the destination-next hoppair (d, j) along with the sijd(t), gijd, wijd, and aijd

values described below:(a) sijd(t) is the pheromone trail concentration

left on a trail ij used as a first hop to desti-nation d at current time t due to all the antsthat have traversed the trail, taking intoconsideration the pheromone evaporation(see Eq. (6)). s is thus a weighted measureof how many times the trail ij was traversedby packets intended to d and is thereby ameasure of the goodness of trail ij.

(b) gijd is the heuristic value of going from j to i.In our mapping, g is a measure of the dis-tance to the destination, distijd, going fromi to d, when using next hop j. We setgijd ¼ 1þ 1

distijd.

(c) wijd 2 [0, 1] is the value of the congestionstatus at node j. If wijd = 1, then, node j isconsidered not congested, and if wijd = 0,then the node j is considered congested.The value of w at a node j is measured asthe ratio of empty space in packets in theIP queue size to the number of packetsalready in the IP queue at j.

(d) We see that the goodness of a next hop j isdirectly proportional to sijd(t), inverselyproportional to distijd and directly propor-tional to wijd. Thus, we write:

aijd ¼ ðcs � sijdðtÞaÞ � ðcg � gbijdÞ � ðcw � wc

ijdÞð1Þ

where cs > 0, cg > 0, and cw > 0 are arbi-trary constants, and a,b,c are integers suchthat a,b,c > 0.

For our use, we need to normalize theabove value of aijd so that we may gaugethe relative effectiveness of each next hop.We normalize it such that aijd 2 [0, 1]:

aijd ¼ðcs � sijdðtÞaÞ � ðcg � gb

ijdÞ � ðcw � wcijdÞ

Pl2J ðcs � sa

ildÞ � ðcg � gbildÞ � ðcw � wc

ildÞð2Þ

where J is the set of next hops at i to desti-nation d. We then set cs = cg = cw = 1, andarrive at

aijdðtÞ ¼½sijdðtÞ�a½gijd �

b½wijd �c

Pl2J ½sildðtÞ�a½gild �

b½wild �c ð3Þ

where a, b and c are chosen appropriately(see Section 4). The above formula was


ARTICLE IN PRESS

adapted from Ant Colony Optimizationtechniques outlined in [18]. The intuitionbehind this equation is that we want touse the metrics of hop distance and pathgoodness and allow some flexibility as tohow much we rely on either metric by vary-ing a, b and c values.

As soon as an ant p is received at a node i vianeighbor node j, i has the information about j’s con-gestion status from Sp. The pheromone sp

ijv depos-ited by an ant p and the heuristic gijv to adestination v in the ant p traversing from node jto node i via nodes v 2 V are given by the equations:

spijv ¼

1

pj � pi

ðv; i; j 2 V Þ ð4Þ

and

gpijv ¼ 1þ 1

depthðvÞ ðv; i; j 2 V Þ ð5Þ

where pi and pj are the pheromone amounts of ant pat nodes i and j, respectively, and V = {v1,v2, . . . ,vm} denotes the set of m nodes visited by p. Thevalue depth(v) is the depth of the node v in p’s nodesvisited stack.All s values in Ai are evaporatedaccording to Eq. (7) each time another ant, p 0, visitsnode i. Let us say p 0 traverses the same trail ij attime (t + D) as traversed by p at time t. p 0 then pos-

itively reinforces the trail ij "v 2 V in Sp. All othertrails iJ 0, (where J 0 is the set of all possible next hopsfrom i except j) in the ant decision table Ai are notpositively reinforced, and in the event no anttraverses through any of the other trails iJ 0, thetrails iJ 0 eventually become invalidated (negatively

reinforced) owing to pheromone evaporation. Thenew sijv at time (t + D) is calculated as follows:

sijvðt þ DÞ ¼ evaporateðsijvðtÞ;DÞ þ sp0

ijv ð6Þ

where sp0ijv is the pheromone deposited on the trail

by p 0 over ij (see Eq. (4)). The function evapo-

rate(sijv(t),D) returns the pheromone amount lefton trail ij for destination v (after evaporation) dueto the ants which traversed ij before p 0. The phero-mone evaporation model used to calculate howmuch of the earlier pheromone trail, sijv(t) is left be-hind at (t + D) when p 0 traverses the trail ij is asfollows:

evaporateðsijvðtÞ;DÞ ¼sijvðtÞ2D=c

ð7Þ

where c is an arbitrary constant. After all the s val-ues in the ant decision table are evaporated (includ-

ing the siJ 0V 0 values on the trails iJ 0 that werenegatively reinforced, i.e., no ants that traveledV 0 ¼ fv01; v02; . . . ; v0m0 g were received) and recalcu-lated, the aijv values for all entries V in Sp0 arerecomputed and the new best next hops to destina-tions V are computed again. This is followed by anupdate of the routing table at node i. Negative rein-forcement of routes also happens when a route isexplicitly invalidated by a route error message.(3) Routing table: The routing table at node i is a

table containing an entry for each destinationd reachable from node i along with the best nexthop, Jd

i , to d. The best next hop, Jdi , to a des-

tination d is the next hop that contains the larg-est aijd value in Ai. The value of Jd

i is therebyupdated every time an ant visits a node i. Therouting table also contains the distance of d

from i in hops, and this information is used toset the number of hops for route discoverywhen the routing table entry to d in i becomesdefunct.In the case of nodes which are part ofhighly capable infrastructure, the routing is sto-

chastic, and the next hop is chosen directly fromthe ant decision table probabilistically. Specifi-cally, a next hop j at node i for destination d

is chosen with a probability of aijd.

3.2.2. Amplification of fluctuations

The process of broadcasting ants during reactive/proactive route discovery/recovery/maintenancefinds new routes to nodes and alters the informationin the ant decision table accordingly. Because of thenature of broadcast in the wireless medium, theroutes found as a result of forward reactive antactivity reflect the current status of the networkand accordingly amplify the current fluctuations inthe topology. Another mechanism amplifies thefluctuations in the local area: when a node receivesa unicast packet, it notes the neighbor node IDand reinforces the path to the source of the packetvia the neighbor. In addition, when a data packetis sent along a next hop, the node reinforces the nexthop as a valid next hop to the destination. Thismechanism also amplifies local fluctuations of net-work and topological characteristics and see to itthat the nodes in the ANSI network use up-to-datenetwork and topological information.

Some protocols, for e.g., [10], using SI mecha-nisms for MANET argue for unicasting forwardreactive ants along one randomly chosen path tothe source and destination to amplify the fluctua-

a b

Fig. 1. Local reinforcement in ANSI. (a) Reinforcement by datapackets. Node i, upon receiving a data packet from S via node j,reinforces the path to node j via j and the source S via j.(b) Reinforcement in neighbor discovery mechanisms. Uponreceiving a HELLO beacon from j all nodes i reinforce trails via j.


ARTICLE IN PRESS

tions in the network. Yet others, for e.g., [16], arguefor sending forward ants at regular intervals fromthe source while the source is sending packets tothe destination. We feel that the above methods willwork in low traffic scenarios and in wired networks,where there is little or no mobility, but not in highlymobile MANET with high loads. Besides, we feelthat the right model for amplification of fluctuationsin a MANET using SI mechanisms is the model weuse: that of broadcasting forward ants only whenabsolutely needed both at the source and intermedi-ate nodes (to perform local repair), and using thesemechanisms with a neighbor discovery mechanism,and applying the rules of SI on the data collected(viz., Eqs. (3)–(7)). By using the SI mechanismsappropriately in ANSI, we are able to reduce thenumber of MAC layer resources used wastefully,as well as be responsive to a network with high traf-fic rates, and provide better packet delivery ratesand lower delay jitter characteristics.

3.3. Protocol description

A trail ij to destination d, sijd, is positively rein-forced in ANSI when (a) a new route to a destina-tion d is found (via ant activity) at i via next hop(neighbor node) j, and (b) when i uses an alreadyknown nexthop node j again to route a packet tod. A trail ij is negatively reinforced when (a) the trailij to destination d is subjected to evaporation (as perEq. (7)), and (b) when next hop node j to d is nolonger available (owing to MAC layer errors, routeerrors, or congestion at j). In the following sections,we describe the various reinforcement mechanismsat work in ANSI.

3.3.1. Local route management—reinforcement by

data packets and the use of neighbor discovery

HELLO messages

Local route management is made possible byreinforcement due to both movement of data pack-ets and an explicit neighbor discovery mechanism.These two concepts are illustrated in Fig. 1.

When a data packet arrives at a node i via aneighbor node p and is sent to the destination alongnext hop j, both the trail to the previous hop, ip, andthe trail to the next hop, ij, are reinforced by the SImechanisms at i.

In addition, nodes in ANSI periodically broad-cast a HELLO message. This message can containa variety of information about the node sendingthe message, such as congestion status. In ANSI,

hello packets are used to perform local route man-agement by positively reinforcing previously knownneighbors and new neighbors. The advantage ofusing this mechanism can be explained as follows:If a direct route to a destination d is known at i

via this process, then a previously known indirectroute to d is less favored than the direct route bythe reinforcement mechanisms in ANSI. Note thatHELLO messages are sent via all available interfacesto facilitate neighbor discovery over all possiblepaths.

3.3.2. Non-local route management and explicit

positive reinforcement

Reactive route discovery is performed by forward

reactive ants, pf, and backward reactive ants, pb.Reactive route discovery can be used both at thesource of a data packet and at an intermediate nodelooking for an alternate route to the destination inthe event that previously known routes to the desti-nation have proved ineffective. A route request issent by deploying a forward reactive ant pf andthe route reply is sent using a backward reactiveant, pb. Even though multiple routes can be gath-ered by a source sending forward reactive ants (byallowing the destination to send backward reactiveants in response to all copies of the forward reactiveants received), we allow the destination to send abackward reactive ant only for the first forwardreactive ant received. This is because we found thatin a high traffic/mobility scenario in which aMANET node has many routes to the destination,packet delivery from source to destination cansuffer invariably because using several routes willspread the traffic over more nodes, and increase

Fig. 2. The propagation of the forward reactive ant (shown insolid arrows) and the return of the backward reactive ant (dashedarrows). The rebroadcast from node 2, when received at node 1 iskilled immediately to prevent route loops. At each node X theforward reactive ant enters, it reinforces the path from X to all theother nodes in the nodes-visited-stack. Thus, the forward reactiveant from S when received at node 4 reinforces the trail 4–2 toboth node 2 and node S. On the return path, all nodes in path ofthe backward reactive ant reinforce the trails to the paths to allthe nodes in the path leading from the node upstream all the wayto the destination. Thus, when the backward reactive ant isreceived at S via path 1–3, . . . ,D, S will reinforce trail S � 1 fordestinations 1,3, . . . ,D.


ARTICLE IN PRESS

the contention in the network. In this case, it seemslike using one route deterministically, while keepingtabs on the congestion status of neighboring nodes(which is what ANSI does) is a better approach.3

Regardless, multiple routes are collected owingto the interaction of the ant information from thenodes in the network and HELLO beacons, andare used as and when older routes become defunct.Also, note that regardless of collecting informationabout multiple routes via other mechanisms, ANSIuses a deterministic choice of next hops when usingpure MANET nodes (highly capable nodes collectmultiple routes and use stochastic routing, as wewill see later). This is because we found that stochas-tic approaches in MANET nodes using ANSI arenot suited to high data delivery in high trafficscenarios.

In Fig. 2, consider a node S which needs to routedata packets to D, but does not have a route to D.Node S buffers the data destined for D and broad-casts (over all interfaces) a forward reactive ant,pSD

f (with a nodes visited stack Sf ), intended to dis-cover the route to D. Because there is a good chancethat D has moved, the current implementation ofANSI sets the number of hops, /f, for the forwardreactive ant (sent from S) to be a few hops largerthan the last known distance of S from D, whichcan be obtained from the routing table at S. If Sreceives data intended to D after pSD

f has beenbroadcast, S buffers the data. When D receivespSD

f , D copies the nodes visited stack, Sf , into anew backward reactive ant, pDS

b , and kills pSDf . D

then sends pDSb to S. pDS

b is not broadcast, it justbacktracks to the source S by using the nodes vis-ited stack Sb in pDS

b . The ant, pDSb , when visiting a

node X along the path to S positively reinforcesthe route to all nodes v 2Sb upstream from X toD, and adds an entry in AX to D via the next hopimmediately upstream (in the path from S to D).An intermediate node thereby knows what nexthop to use to route to D. In this way, backwardreactive ants perform explicit positive reinforcementof routes to destination D. When S receives pDS

b

from D, S sends the buffered packets intended forD over the newly discovered route and flushes S’sbuffer. Note that multiple paths may be readily col-lected (for example, by sending another backwardreactive ant for the ant proceeding to D via nodes

3 Stochastic approaches to routing in pure MANET networksis an effective approach when the mobility and traffic in thenetwork are low.

2 and 4), but this is not done for pure MANETnetworks.

In the event that pDSb is not received at S within a

timeout period, then the value of /f is increased by 2more hops and the search for the route resumesagain. The process of route discovery is continuedagain if a route is not found after the second try.ANSI retries twice for a route to destination.

To control the amount of MAC layer usage at anode X, a scheduled HELLO packet is broadcast atX only if the last broadcast forward reactive antwas sent before the last HELLO message.

3.3.3. Route errors, and negative reinforcement

Route errors occur at a node X when X is unableto provide a route for the destination D owing tonon-availability of a routing table entry at X ordue to the non-availability of the next hop suggestedby the routing table entry at X. When a route erroroccurs at a node X in a network running ANSI, X

first buffers the packet which X needs to forwardand then sends a forward reactive ant to find thedestination D. If X happens to be an intermediatenode, in addition to sending a forward reactiveant, X also sends a route error back to the sourceS of the packet. The packets buffered at X arerelayed across the network after a backward reac-tive ant from D reaches X.

Fig. 3. In a hybrid network (nodes 1–9 are part of a highlycapable network in this case, and are connected by the gridshown), nodes belonging to more capable infrastructure are ableto perform stochastic routing. In this figure, two possible pathsthat 1 can take to route to D are one via nodes 1! 4! 7!8! 9 (P1) and another via nodes 1! 2! 5! 6! 9 (P2).


ARTICLE IN PRESS

In addition, when a route error is received at anintermediate node between X and S, the nodeexplicitly invalidates the routing table entries to D.The packets received at X before the route error isreceived at S are X’s responsibility (to forward),but the packets generated after the time when theroute error is received at S from X are S’s responsi-bility—S generates a forward reactive ant to find theroute to D.

3.3.4. Proactive routing within highly capable sections

of the networkAs mentioned in Section 3.1, nodes belonging to

non-mobile, highly capable infrastructure, such ascellular networks engage in proactive routing as wellas reactive routing because these nodes are not con-cerned about topological fluctuations. These nodesalso maintain a list of mobile nodes which are acces-sible from each other, thus assisting the reactiverouting process within the mobile nodes as andwhen possible. Nodes in non-mobile, highly capableinfrastructure send proactive ants periodically to allthe other highly capable nodes they are connectedto. Proactive ants are not returned like forwardreactive ants, and they reinforce the route to theproactive ant sender along the path the proactiveant takes. Proactive ants, apart from carrying anodes-visited-stack for gathering information aboutthe nodes that were visited, are fixed in hop lengthand also carry a data structure for indicating themobile nodes which are accessible from the proac-tive ant sender. These nodes engage in proactiveroute collecting activity using all their interfaces,and so are able to combine routes found via differ-ent interfaces effectively during the routing process.

Because nodes belonging to a highly capable net-work need not be concerned about the issues due tomobility, these nodes are able to effectively utilizethe benefits due to stochastic routing (see Fig. 3).As mentioned before, ant-based routing naturallylends itself to stochastic routing because multipleroutes are found and maintained.

3.3.5. Driving the routing process via more desirablenodes

By choosing higher values for a, b, and c, theprocess of next hop selection in ANSI favors thenext hops with higher values for s, g, and w, respec-tively. However, by choosing values which are toohigh, the route selection is too skewed towards thebest next hops and it becomes very difficult for theSI mechanisms at the nodes to respond quickly to

the changes in the network. Hence, a choice for a,b, and c should be made carefully to allow forresponsiveness of the system. Using insights fromour preliminary results, we arrived at a value ofa = b = c = 2, and these are the values we use inour implementation.

4. Simulation results

ANSI was simulated in QualNet (Version 3.7),and the performance of ANSI was compared witha popular routing protocol, AODV [1], for the samenetwork and load characteristics. We chose to com-pare ANSI with AODV because AODV has beenshown to perform well in a vast majority of adhoc network scenarios. In addition, AODV alsoworks on hybrid ad hoc networks. Our work hereis an extension of our earlier work [19] which onlytested ANSI under UDP loads over a pureMANET. As we mentioned earlier, ANSI functionsas a purely reactive protocol in a pure MANETenvironment.

4.1. Simulation and network model

4.1.1. ANSI parametersThe current implementation of ANSI used

a = b = c = 2. In both AODV and ANSI, the reac-tive route recovery is retried twice, and for ANSI,the last try uses /f = 15. For the first two trials inANSI, /f is determined according to the informa-tion available about the unknown destination: ifthe destination had a valid entry in the routing tableearlier, then /f is set to one more than the earlier

a b

Fig. 4. Hybrid network topologies used for Experiments 1–3: (a) The hybrid network topology for Experiments 1 and 2. (b) The hybridnetwork topology for Experiment 3.

4 Node density is defined as the number of nodes in an areacovered by the transmission range of a node.


ARTICLE IN PRESS

number of hops to the destination. Otherwise,/f = 5. The evaporation constant, c, used in Eq.(7) is 15 s. In hybrid networks, the nodes whichare part of the high-speed Ethernet (see Section4.1.2) used a proactive route update interval of 10 s.

4.1.2. Network and application parameters

We performed five experiments, in which westudied the performance of ANSI and AODV withincreasing traffic and increasing number of nodesof both UDP and TCP flows in both hybrid andpure MANETs. In all these experiments, the sourceand destination are chosen randomly and are pair-wise-distinct for each trial.

In Experiments 1 and 2, we studied the perfor-mance of ANSI vs. AODV in a hybrid network,for both UDP (Experiment 1) and TCP (Experiment2) flows. In these experiments, the non-mobile nodesare connected to each other over a 100 Mbps Ether-net link. Fig. 4(a) shows the simulation topology.The size of the entire terrain is 2000 m · 2000 m.Inside this terrain, there are four MANET‘‘regions’’, each of which contain 20 MANET nodesinside a terrain of size 500 m · 500 m, and ‘‘ser-viced’’ by one highly capable, immobile node (nodes81–84) located in the center of the mobile region.This highly capable node, located in the center ofeach of the regions, manages both an Ethernet inter-face and an 802.11 interface, and is connected to theothers by another highly capable node, node 85,which has 4 Ethernet interfaces. Note that MANETnodes within a region are not able to communicatewith MANET nodes of other regions directly (theclosest they can get is around 353 m, which is beyondthe transmission range of the MANET nodes).

Four streams are chosen for each region, withone stream headed towards each of the regions

(thus, one stream will be an ‘‘internal’’ stream).There are thus, altogether, 16 traffic streams in thisexperiment. Each of these streams send 512-bytepackets at a uniform rate of 1–20 packets/s.

In Experiment 3, we studied the performance ofANSI and AODV in a larger hybrid network con-sisting of 360 pure MANET nodes spread over 9mobile regions uniformly located in a 5000 m ·5000 m terrain, each of size 1000 m · 1000 m andserviced by one highly capable, immobile node(nodes 361–369) located at the center of each mobileregion. The highly capable nodes are all connectedvia a 100 Mbps Ethernet link. The topology of ourexperiment is shown in Fig. 4(b). Each highly capa-ble node has both Ethernet interfaces and an 802.11interface. The size of the data packets sent was512 bytes. Six UDP streams are randomly gener-ated, with the following profile of the source–desti-nation pairs: (a) regions 1–4, (b) regions 1–7, (c)regions 8–5, (d) regions 8–2, (e) regions 3–6, and(f) regions 3–9. The data sources generated packetsat the uniform rate of 2 to 20 packets/s in steps of2 packets/s.

In Experiment 4, we studied the effect of increas-ing TCP traffic in a pure MANET network. In thisexperiment, 50 nodes were placed uniformly in anetwork of size 1100 m · 1100 m. This maintains anode density4 of 8.15 m�2, which, according to[20], is sparse for a network with mobile nodes.The experiment simulates 25 streams of TCP trafficsending 64-byte packets at a uniform packet ratevarying from 1 to 20 packets/s.


ARTICLE IN PRESS

In Experiment 5, we studied the performance ofANSI and AODV under UDP loads in a pureMANET environment with an increasing numberof nodes. The number of nodes was varied from50 to 250 and exactly half the number of nodes weredata sources. The terrain size was such that the nodedensity was constant at 8.15 m�2 (for example, for50 nodes, the terrain size was 1100 m · 1100 m).The data sources generated one 64-byte packet asecond to be sent to the data sink.

In all the experiments, the MANET nodes wereuniformly distributed initially in the terrain andthe mobile nodes moved as per the Random Way-point Model with a minimum speed of 0.001 m/s,maximum speed of 20 m/s, with a pause time of10 s. In hybrid networks (Experiments 1–3), themobile nodes were restricted to move only withintheir region (bounded by a 500 m · 500 m terrainfor Experiments 1 and 2 and a 1000 m · 1000 m ter-rain for Experiment 3). The MANET nodes in theexperiments used one 802.11 interface with omnidi-rectional antennas and a transmission range of250 m at the physical layer and 802.11DCF at theMAC layer. The link bandwidth for the mobilenodes using 802.11 was 2 Mbps. In addition to using802.11, the non-mobile nodes also used Ethernetwith a capacity of 100 Mbps. The simulations useda two-ray pathloss model and no propagation fad-ing model was assumed. The application used wasCBR, and sources and destinations were pairwisedistinct and chosen randomly. Both TCP andUDP-based CBR flows were studied. Super applica-tion was used for generating a reliable CBR trafficstream using TCP (regular CBR application usesUDP).

All experiments were run for a simulated time of5 min and all sources started sending packets atexactly 40 s into the simulation and ended data gen-eration at exactly 260 s. TCP-LITE, used for theTCP flows in our experiments, is a variant ofTCP-RENO, and used an MSS of 512 bytes, maxi-mum send/receive buffer of 16384 bytes each, anddelayed ACKs.

We studied the following end-to-end and net-work-wide characteristics:

1. (End-to-end metric 1) Packet delivery fraction:

measured at the application layer as the ratio ofthe total number of packets which were received(at the application layer) at the data sinks to thetotal number of packets that were sent from thedata sources (at the application layer), and aver-

aged over the number of source–destinationpairs. For TCP flows, the above described quan-tity is the measured packet delivery ratio. Theactual packet delivery for TCP flows is calculatedusing the expected number of packets that shouldbe sent at the application layer at the datasources. UDP does not perform congestion-con-trol so the expected and measured number ofpackets sent at the application layer of the datasource are the same.

2. (End-to-end metric 2) End-to-end delay: measuredas the average delay in sending packets fromsource to destination and averaged over the num-ber of source–destination pairs.

3. (End-to-end metric 3) Delay jitter: measured asthe average variance of the interarrival times atthe destinations and averaged over the numberof source–destination pairs.

4. (End-to-end metric 4) Number of packets sent by

Super application sender: measured as the totalnumber of packets which are actually sent bySuper Application senders. For Super Applica-tion using TCP, this number depends on howlong the TCP connection lasts.

5. (End-to-end metric 5) Variation of the congestion

window of a sender: measured as the TCP conges-tion window (snd_cwnd) at one sender for oneflow for one trial as it varies with simulation time.

6. (Network-wide metric 1) Total number of route

errors initiated: is the total number of routeerrors generated in the network.

7. (Network-wide metric 2) Total number of 802.11

DCF MAC layer unicasts sent: is the total num-ber of all (successful) 802.11DCF unicast trans-missions sent in the network. For AODV, thismeasures the total number of data packets,RREP and RERR sent out at the 802.11DCFinterface. For ANSI, this measures the totalnumber of data packets, backward reactive ants,and RERRs sent at the 802.11 interface.

8. (Network-wide metric 3) Total number of 802.11

DCF MAC layer broadcasts sent: is the totalnumber of all 802.11DCF broadcasts sent by allnodes in the network. For AODV, this measuresthe total number of RREQ and Hello packetssent at the 802.11DCF interfaces, and for ANSI,this measures the total number of forward reac-tive ants, proactive ants and the Hello packetssent at the 802.11 interfaces.

We do not report end-to-end delay and delay jit-ter for TCP flows as these metrics are typically not


ARTICLE IN PRESS

reported for TCP flows, because of the fact thatTCP has to deal with out-of-order deliveriesand the large delays (as compared to UDP flows)owing to reliability and congestion-control mecha-nisms.

We analyzed the results from the above experi-ments and show them using graphs with 95% confi-dence intervals of the measured values.

4.2. Simulation results

4.2.1. Experiment 1: Hybrid network—effect ofincreasing the UDP packet rate

Fig. 5 shows the results for the performance ofANSI vs. AODV over a hybrid network usingUDP flows. We see that ANSI consistently outper-forms AODV in terms of packet delivery, delay, jit-ter and number of RERR initiated. ANSI and

0 5 10 15 200.95

0.96

0.97

0.98

0.99

1

Packet rate (pkts/s)

Pac

ket d

eliv

ery

at th

e C

BR

laye

r (%

)

AODV

ANSI

0 5 10 15 200

0.05

0.1


Del

ay ji

tter

(s)

AODV

ANSI

0 5 10 15 200

5

10

15x 10

4


802.

11D

CF

,Uni

cast

s se

nt

AODV

ANSI

a b

c

e f

Fig. 5. Experiment 1: Performance studies of ANSI vs. AODV in a hybend delay, (c) delay jitter, (d) number of RERR initiated, (e) 802.11DC

AODV send a comparable number of MAC uni-casts. ANSI sends fewer MAC broadcasts whenthe packet rate is low to moderate, but as the packetrate increases, ANSI sends more MAC broadcasts.

The reason why ANSI performs better thanAODV—delivering more packets with better met-rics such as delay, jitter and number of routeerrors—is because ANSI manages the local networkinformation better than AODV does, and performscongestion-aware routing. This is why ANSI showslower route errors as compared to AODV (seeFig. 5(d)). Owing to the above reasons, routes breakless often and result in fewer route request opera-tions in ANSI as compared to AODV. When routesdo break in ANSI, they are managed by the proto-col mechanisms locally rather than a network-wideflooding. This in turn results in lower congestionat the nodes. This is why, even though ANSI shows

0 5 10 15 200

0.05

0.1

0.15

0.2

0.25

0.3


End

-to

end

dela

y (s

)

AODV

ANSI

0 5 10 15 200

100

200

300

400

500

600


Tot

al n

umbe

r of

RE

RR

initi

ated

AODV

ANSI

0 5 10 15 200

1

2

3

4

5x 10

4


802.

11D

CF

,Bro

adca

sts

sent

AODV

ANSI

d

rid network with UDP flows: (a) Packet delivery ratio, (b) end-to-F, Unicasts sent and (f) 802.11DCF, Broadcasts sent.

5 This amount is 16 · 220 · x = 3520x packets total (x is thepacket rate), and indicated by the straight line graph in Fig. 6(b).


ARTICLE IN PRESS

larger MAC broadcasts in higher packet rates, itstill shows delays and jitter lower than that forAODV. The higher number of broadcasts whenthe packet rate increases is because of the conges-tion-aware properties of ANSI, which allow it todrop badly congested routes and look for new ones.This, while delivering packets more quickly andsmoothly, obviously makes ANSI incur more routediscovery overheads, which is what we see in termsof larger MAC broadcast overheads. The new, con-gestion-free (or low congestion) routes are then usedto deliver more packets in ANSI. Note that AODV,does not show an appreciable increase in the num-ber of MAC broadcasts as the packet rate increasesbecause it does not perform congestion-aware rout-ing, but owing to this, the performance of AODVdegrades. The fact that the number of ANSI’sMAC unicasts are comparable to that of AODV(in the context of better performance metrics), alongwith its fewer route errors is a clear indication of thefact that ANSI is engaged in providing/finding bet-ter routes as compared to AODV.

The reason why delay jitter decreases withincreasing packet rate in both ANSI and AODV(see Fig. 5(b) and (c)) is as follows: Delay jitter isa measure of the variation of interarrival times atthe destination. Thus, if end-to-end delay measuredat the destination varies very little, then delay jitteris bound to be low. ANSI, being congestion-aware,chooses congestion-free routes and delivers packetsat the destination with little variation in end-to-end delay. AODV, because it is not congestion-aware, delivers packets along congested routes,which results in higher end-to-end delays becausea node running AODV does not react to congestionuntil a congested node along the path is no longerable to receive or transmit packets. Thus, for a sin-gle stream of UDP traffic from one source to desti-nation in AODV, the destination first experienceslow variation in end-to-end delay, but thereafter,the path becomes more congested and the variationin end-to-end delay progressively increases until thepath breaks. AODV then engages in route discoveryand finds a congestion-free path, and once again themeasurement of end-to-end delay at the destinationshows low variation until the new path becomescongested again. Statistically, the value of delay jit-ter depends on the percentage of the packets thatare delivered at the destination with low variationin end-to-end delay, and so if a higher percentageof packets are delivered with a higher variation,the jitter is bound to be larger.

As packet rate increases, for both AODV andANSI, the mean time before links break owing tonode mobility is still the same, but because the ofthe use of highly capable nodes (which are within2 hops away for any MANET node), the percentageof packets delivered with lower variation in end-to-end delay increases (in comparison to the number ofpackets delivered at higher variations in end-to-enddelay) at both AODV and ANSI, thus bringing theoverall variation down. This is why we see adecrease in delay jitter as packet rate increases forboth ANSI and AODV.

4.2.2. Experiment 2: Hybrid network—effect of

increasing the TCP packet rate

Fig. 6 shows the results for Experiment 2. ForTCP flows, we see that ANSI’s measured and actualpacket delivery ratio is higher than the same metricsfor AODV. We also see that for AODV, the mea-sured packet delivery ratio improves as the packetrate increases, but the actual packet delivery ratiodecreases. ANSI’s actual packet delivery is nearly5–10% more than AODV’s actual packet deliveryratio. ANSI also sends more packets during the sim-ulation as compared to AODV—we see that thenumber of packets which ANSI sends is very closeto the number expected to be sent.5 In terms ofthe effect of the routing protocol on TCP, the con-gestion window for the output queue at node 53(for packet rate 1 packets/s, sent from node 53 tonode 48) shows steady growth, while AODV’s con-gestion window (for the same stream, output queueat node 53) shows substantially slower growth.ANSI, as before, shows a lot fewer route errors(see Fig. 6(d)). ANSI shows more MAC unicasttraffic as compared to AODV. Though ANSI showslower MAC broadcast traffic when the packet rate islow, it shows more MAC broadcast overheads whenthe packet rate increases.

The reason why ANSI performs better (with 5–10%higher actual packet delivery ratio) than AODV underTCP loads is because of the congestion-aware routingin ANSI. Owing to this property, ANSI is able to sup-ply congestion-free routes which allow for the smoothpassage of ACKs back to the data source, allowingTCP operations to perform smoothly.

We would like to draw attention to the graphsshowing the measured packet delivery ratio inFig. 6(a). These results for measured packet delivery

0 5 10 15 200.8

0.85

0.9

0.95

1


Pac

ket d

eliv

ery

at th

e ap

p. la

yer

(%)

AODV(m)AODV(a)ANSI(m)ANSI(a)

0 5 10 15 200

2

4

6

x 104


Pac

kets

sen

t by

the

app.

laye

r

AODV

ANSI

Ideal

0 100 200 3000

5000

10000

15000

Simulation time (s)

Con

gest

ion

win

dow

(by

tes) ANSI

AODV

0 5 10 15 200

50

100

150

200

250


Tot

al n

umbe

r of

RE

RR

initi

ated

AODV

ANSI

0 5 10 15 200

0.5

1

1.5

2

2.5

3x 10

5


802.

11D

CF

,Uni

cast

s se

nt

AODV

ANSI

0 5 10 15 200

1

2

3

4x 10

4


802.

11D

CF

,Bro

adca

sts

sent

AODV

ANSI

a b

c d

e f

Fig. 6. Experiment 2: Performance studies of ANSI vs. AODV in a hybrid network with TCP flows: (a) measured (m) and actual(a) packet delivery ratio, (b) packets sent by the Super Application Layer, (c) congestion window for TCP output at node 53 for 1 pkt/s,(d) number of RERR initiated, (e) 802.11DCF, Unicasts sent and (f) 802.11DCF, Broadcasts sent.


ARTICLE IN PRESS

ratio are counter-intuitive. While the measuredpacket delivery ratio of AODV increases withpacket rate, we note that the percentage of packetssent to the data sink increasingly decreases. Thus,the actual packet delivery ratio, measured as thepercentage of packets that are received to the per-centage of packets that are expected to be sent (inthis case x · 16 · 220 = 3520x, where x is the packetrate), actually decreases. So, the traditional packetdelivery ratio metrics, defined as the ratio of thenumber of application layer packets delivered tothe number of application packets sent, is actuallya misleading metric to measure when studyingMANET performance under TCP loads.

The behavior of ANSI and AODV under TCPloads can be summarized clearly by Fig. 6(c). Note

that we had fixed the TCP send buffer to be16,384 bytes, and the congestion window cannotgrow beyond this size. In this figure, we see howSuper application works TCP when sending CBRtraffic. Note that this is traffic inside a mobile region(both node 53 and node 48 are inside the samemobile region as per Fig. 4(a)). TCP, when workingon top of ANSI, is able to increase the congestionwindow as per congestion avoidance algorithms,but in AODV, congestion avoidance is quicklythwarted by congestion occurring along the pathfrom node 53 to node 48, which is why the TCP stackat node 53 shows fast recovery behavior [21] for theTCP output queue. This is the case owing to losing alot of ACKs in AODV. Indeed, we see that the con-gestion window in AODV does not grow/change


ARTICLE IN PRESS

after a certain point into the simulation (around200 s) for AODV. Whereas, for ANSI, we see a‘‘healthy’’ growth of the congestion window, con-trolled by the congestion-avoiding sender (lineargrowth of congestion window) rather than beingcontrolled by congestion elsewhere in the network.

This behavior for TCP running over ANSIresults from ANSI’s congestion awareness, whichconstantly maintains routes with low congestionand chooses them in favor of the ones with highercongestion. This permits TCP running over ANSIto receive ACKs more frequently and regularly thanin the AODV case, where losing ACKs causes fastrecovery behavior. AODV, not being congestionaware, chooses congested routes frequently becauseit has no way of knowing which routes are con-gested and which ones are not, making the passageof ACKs more difficult.

5 10 15 200.7

0.75

0.8

0.85

0.9

0.95

1


Pac

ket d

eliv

ery

at th

e C

BR

laye

r (%

)

AODV

ANSI

0 5 10 15 200

0.1

0.2

0.3

0.4


Del

ay ji

tter

(s)

AODV

ANSI

0 5 10 15 200

2

4

6

8

10x 10

4


802.

11D

CF

,Uni

cast

s se

nt

AODV

ANSI

a b

c d

e f

Fig. 7. Experiment 3: Performance studies of ANSI vs. AODV in a (la(b) end-to-end delay, (c) delay jitter, (d) number of RERR initiated, (e

More packets are sent by ANSI at the Superapplication layer as a result of larger congestionwindows, and subsequently, more MAC unicastsare sent for these packets, which is why we see thenumber of MAC unicasts for ANSI is more. MoreMAC broadcasts are sent in ANSI as a responseto finding newer routes which are less congested.As before, AODV does not respond to congestion,and so it shows only a small increase in the numberof MAC broadcasts as the packet rate increases.

4.3. Experiment 3: Large hybrid network—effectof increasing UDP packet rate

Fig. 7 shows the results for the performance ofANSI and AODV in a larger hybrid network. Aswe can see, the results are similar to the results ofExperiment 1, shown in Fig. 5. We also see ANSI’s

0 5 10 15 200

0.1

0.2

0.3

0.4


End

-to-

end

dela

y (s

)

AODV

ANSI

5 10 15 200

500

1000

1500


Tot

al n

umbe

r of

RE

RR

initi

ated

AODV

ANSI

0 5 10 15 200

0.5

1

1.5

2

2.5

3

x 105


802.

11D

CF

,Bro

adca

sts

sent

AODV

ANSI

rger) hybrid network with UDP flows: (a) packet delivery ratio,) 802.11DCF, Unicasts sent and (f) 802.11DCF, Broadcasts sent.


ARTICLE IN PRESS

packet delivery metrics are on average around 10%more; ANSI also delivers these packets in roughly1/3 as much time as AODV, with lower delay jitter,and fewer route errors. As with Experiment 1, wealso see that these performance improvements aremade in ANSI at the cost of higher MAC layerresource consumption owing to the congestion-aware routing in ANSI.

ANSI’s performance in comparison to AODV isbetter in Experiment 3 than in Experiment 1 isbecause of the fact that in Experiment 3, ANSIis able to take advantage of proactive routing/stochastic routing in the highly capable nodes.

The reason why packet delivery decreases forboth ANSI and AODV more drastically (as packetrate increases) for this experiment as compared toExperiment 1 is because of the effect of a larger

0 5 10 15 200.4

0.5

0.6

0.7

0.8

0.9

1


Pac

ket d

eliv

ery

at th

e ap

p. la

yer

(%)

AODV(m)

AODV(a)

ANSI(m)

ANSI(a)

0 100 200 3000

5000

10000

15000

Simulation time (s)

Con

gest

ion

win

dow

(by

tes) ANSI

AODV

0 5 10 15 200

2

4

6

8

10

12x 10

4


802.

11D

CF

,Uni

cast

s se

nt

AODV

ANSI

a

c

e

Fig. 8. Experiment 4: Performance studies of ANSI vs. AODV in a puredelivery ratio, (b) packets sent by the Super Application Layer, (c)(d) number of RERR initiated, (e) 802.11DCF, Unicasts sent and (f) 8

mobile area. In Experiment 1, the highly capablenode in each mobile region is in the worst 353 maway (250 m�

ffiffiffi2p

) from a mobile node, which is2 hops, but in Experiment 3, the highly capablenode is in the worst case 707 m away, which is 3hops away. Thus, the effects of traffic under highmobility scenarios weigh in more in Experiment 3than in Experiment 1.

Finally, we note that the graphs for end-to-enddelay, Fig. 7(b), and delay jitter, Fig. 7(c), are simi-lar to the corresponding graphs for Experiment 1for the same reasons.

4.3.1. Experiment 4: Pure MANET—effect

of increasing the TCP packet rate

In [19], we showed that the ANSI network is ableto do better than the AODV network in a pure

0 5 10 15 200

1

2

3

4

5

6x 10

4


Pac

kets

sen

t by

the

app.

laye

r

AODV

ANSI

Ideal

0 5 10 15 200

100

200

300

400

500

600

700


Tot

al n

umbe

r of

RE

RR

initi

ated

AODV

ANSI

0 5 10 15 200

1

2

3

4

5

6

7x 10

4


802.

11D

CF

,Bro

adca

sts

sent

AODV

ANSI

b

d

f

MANET with TCP flows: (a) measured (m) and actual (a) packetcongestion window for TCP output at node 5 for 1 packets/s,02.11DCF, Broadcasts sent.


ARTICLE IN PRESS

MANET scenario for UDP flows. Here, we showthat the ANSI network is able to do better thanthe AODV network for TCP flows as well. Fig. 8shows the performance of ANSI and AODV overa pure MANET for TCP flows. We see that ANSIis able to deliver more packets at a higher packetdelivery ratio (both measured as well as actual) ascompared to AODV, and in Fig. 8(c), we see thatTCP over ANSI is able to increase its congestionwindow, whereas TCP over AODV is not able todo the same. We also see that ANSI shows fewerroute errors as compared to AODV. In terms ofMAC layer overheads, ANSI shows higher over-heads, both with unicast and broadcast MACtraffic.

The packet delivery ratio (measured and actual)and number of packets sent metrics are better forANSI because it manages its routes better and alsoperforms congestion-aware routing. As far aspacket delivery ratio and number of packets sent(see Fig. 8(a) and (b)) is concerned, in Fig. 8(b),we see that the total number of packets sent bythe application for either protocol is actually wellbelow the total number which is expected to be sent,under ideal cases.6 So, we see that the actual packetdelivery ratio for either protocol decreases, eventhough the measured packet delivery ratio is veryhigh. We note however that ANSI is closer to theexpected number of packets sent as compared toAODV. As with Experiment 2, we note that study-ing measured packet delivery ratios for MANETsunder TCP loads is misleading without Fig. 8(b).

In Fig. 8(c), we see that ANSI’s effect on TCP isbetter than AODV’s effect on TCP—indeed, we seethat the AODV congestion window drops to512 bytes (MTU of the 802.11 interface) and staysthere after (about) only 1/4 of the simulation time.We also see, as before, that AODV generates a lotmore route errors as compared to ANSI because ithas no way of telling which routes are headingtowards congestion and which ones are not, makingroutes break more often. The ANSI network sendsmore unicast packets as compared to AODVbecause more packets are delivered in the ANSI net-work as compared to the AODV network. Thenumber of broadcasts are also higher in the ANSInetwork. As before, this is owing to the conges-tion-aware routing performed by ANSI.

6 This amount is 25 · 220 · x = 5500x (x is the packet rate), asindicated by the straight line in Fig. 8(b).

4.4. Experiment 5: Pure MANET—effect of

increasing the number of nodes under UDP flows

Fig. 9 shows the results for Experiment 5. Wesee that ANSI has consistently better (or compara-ble) packet delivery metrics, fewer RERR, andfewer MAC broadcasts as compared to AODV.We also see that the end-to-end delay, delay jitter,and MAC unicasts for ANSI are more thanAODV until about 150 nodes, after which the met-rics for ANSI are better or comparable to that ofAODV.

We explain the above results as follows: when thenumber of nodes and the terrain size increases(recall that the terrain size was increased in thisexperiment to keep node density constant), the aver-age length of the route increases in ANSI owing toits congestion-aware routing. In addition, half thenumber of nodes in the network are data sources,making the effect of congestion in the network animportant factor in deciding routes. This is whywe see that ANSI has larger delays (owing to longerpaths) with higher delay jitter and more MAC uni-casts when the number of nodes increases, until 150nodes. Under these conditions, the expenses of con-gestion-aware routing are more than the expenses ofnot performing the same. Note that owing to theabsence of highly capable nodes, the effects due tolink breakages in the network only increase as thenetwork increases in size. This is why end-to-enddelays and delay jitter increase as the number ofnodes increase.

As the number of nodes in the network increasebeyond 150 nodes and the traffic increases, theeffects due to bad management of routes in AODVfar outweigh the advantages due to sending packetsvia the shortest paths, which creates hotspots. Thisis why we see ANSI’s performance degradation ismore gradual and steady, whereas AODV’s deterio-ration in performance metrics is unstable and steep.This insight is corroborated in Fig. 9(d) and (f),where we see that the AODV loses more routes thanANSI does, and thereby engages in route discoveryactivity more often (as seen by the amount of MAClayer broadcasts sent).

4.5. Discussion

ANSI is able to perform better as compared toAODV owing to a combination of both better routemanagement and congestion-aware characteristics.In addition, in hybrid ad hoc networks, ANSI is

50 100 150 200 2500

0.2

0.4

0.6

0.8

1

Number of nodes

Pac

ket d

eliv

ery

at th

e C

BR

laye

r (%

)

AODV

ANSI

50 100 150 200 2500

0.5

1

1.5

2

2.5

Number of nodes

End

-to

end

dela

y (s

)

AODV

ANSI

50 100 150 200 2500

2

4

6

8

10

Number of nodes

Del

ay ji

tter

(s)

AODV

ANSI

50 100 150 200 2500

2000

4000

6000

8000

10000

Number of nodes

Tot

al n

umbe

r of

RE

RR

initi

ated

AODV

ANSI

50 100 150 200 2500

2

4

6

8

10x 10

4

Number of nodes

802.

11D

CF

,Uni

cast

s se

nt

AODV

ANSI

50 100 150 200 2500

1

2

3

4

5x 10

6

Number of nodes

802.

11D

CF

,Bro

adca

sts

sent

AODV

ANSI

a b

c d

e f

Fig. 9. Experiment 5: Performance studies of ANSI vs. AODV in a pure MANET network with UDP flows and a varying number ofnodes: (a) packet delivery ratio, (b) end-to-end delay, (c) delay jitter, (d) number of RERR initiated, (e) 802.11DCF, Unicasts sent and(f) 802.11DCF, Broadcasts sent.


ARTICLE IN PRESS

able to harness the power of proactive/stochasticrouting in immobile, highly capable nodes con-nected to each other over Ethernet links. An insightwe gain in the experiments above is with regard tocongestion-aware routing. We see that there arebenefits to doing congestion-aware routing, but itcomes at a cost—MAC layer resources. ANSI,being congestion-aware, results in fewer routeerrors, but incurs a larger number of MACresources in general. Congestion-aware routing pro-tocols will have to invalidate routes more often thantheir counterparts which are not congestion-aware.This presents a trade-off between performance andfinding congestion-free paths. When the traffic loadincreases, which is when performing congestion-aware routing is most useful, this is a very delicate

balance. On the one hand, performing congestion-aware routing adapts the network to congestion,but it also decreases the resources available to senddata. ANSI is able to reach this balance and thushas significant advantages in low traffic networks,as also in high traffic scenarios.

Another insight is regarding the use of TCP tostudy MANET routing protocols. Typically, useof TCP over MANETs is a very divisive idea, witha lot of research leaning towards the fact that it isa bad protocol to use over MANET [22]. Regardlessof the stand taken by MANET transport layerresearchers, we see that studying TCP loads overrouting protocols can lead to better understandingof routing protocols. Here, we see that when reli-ability mechanisms of the routing protocol are used


ARTICLE IN PRESS

along with a transport layer mechanism, the resultscan be very compelling.

5. Conclusions and future work

This paper describes the design, implementation,and performance of a swarm intelligence-basedhybrid routing protocol, ANSI, for hybrid ad hocnetworks. We simulated ANSI and carried out a per-formance comparison with AODV, and the resultsshow that ANSI performs better than AODV forboth UDP and TCP flows in both pure MANETand hybrid ad hoc networks with respect to packetdelivery, number of packets delivered, end-to-enddelay, and delay jitter. We also see ANSI affects theupper layer protocols such as TCP and Super Appli-cation favorably. In addition, we see that the varianceof the observed values (as measured by the width ofthe confidence intervals) is most often lower in ANSI,indicating a more stable performance.

We also see that implementing congestion-aware-ness at the routing layer comes at a cost. In the caseof the ANSI network, congestion awareness trans-lates to higher route discovery activity because therouting layer invalidates congested routes, allowingTCP to perform more smoothly. Even though con-gestion-awareness improves the performance of thenetwork, it does raise scalability issues when the traf-fic increases. In general, we note that it is very diffi-cult to design routing protocols which are scalableunder extreme traffic conditions, but incorporatingcongestion awareness complicates the problem byincurring overheads in an already bogged network.The trade-off between the amount of overheadexpended in finding congestion-free paths and theamount of resources remaining for actual data deliv-ery is very delicate at high traffic loads and is worthstudying. However, we see that ANSI is able toachieve this balance and perform better than AODVin higher traffic load conditions, even though thesecome at the cost of network resources.

We believe that the basic ANSI structure pro-vides for implementing a more general purpose,self-organizing routing protocol incorporatingautonomously adaptive characteristics that enableit to behave well under all network and traffic con-ditions. For example, the nodes-visited-stack usedin the ants in ANSI can be used to collect a widevariety of information at the nodes the ant visits,such as energy reserves at a node, which in turncan be used to make better next hop selections.

References

[1] C.E. Perkins, E.M. Royer, S.R. Das, Ad hoc on-demanddistance vector (AODV) routing, Internet draft (draft-ietf-manet-aodv-09.txt), November 2001, in preparation.

[2] D.B. Johnson, D.A. Maltz, Y.-C. Hu, J.G. Jetcheva, Thedynamic source routing protocol for mobile ad hoc networks(DSR), Internet draft (draft-ietf-manet-dsr-07.txt), February21, 2002.

[3] C. Perkins, P. Bhagwat, Highly dynamic destination-sequenced distance-vector routing (DSDV) for mobilecomputers, ACM SIGCOMM’94 Conference on Communi-cations Architectures, Protocols and Applications (1994)234–244. Available from: citeseer.nj.nec.com/perkins94highly.html [Online].

[4] Z.J. Haas, M.R. Pearlman, P. Samar, The zone routingprotocol (ZRP) for ad hoc networks, July 2002, iETFInternet draft, draft-ietf-manet-zone-zrp-04.txt.

[5] C. Adjih, T. Clausen, P. Jacquet, A. Laouiti, P. Minet, P.Muhlethaler, A. Qayyum, L. Viennot, Optimized link staterouting protocol, INRIA Rocquencourt, Internet draft(draft-ietf-manet-olsr-09.txt), April 15, 2003, in preparation.

[6] V. Ramasubramaniam, Z.J. Haas, E.G. Sirer, SHARP: Ahybrid adaptive routing protocol for mobile ad hoc net-works, in: The ACM Symposium on Mobile Adhoc Net-working and Computing (MobiHoc 2003), Annapolis,Maryland, June 1–3, 2003.

[7] E. Bonabeau, M. Dorigo, G. Theraulaz, Swarm Intelli-gence—From Natural to Artificial Systems, Oxford Univer-sity Press, New York, 1999.

[8] C. Prehofer, C. Bettstetter, Self-organization in communica-tion networks: principles and design paradigms, IEEECommun. Mag. (2005) 78–85.

[9] J.P. Sterbenz, R. Krishnan, R.R. Hain, A.W. Jackson, D.Levin, R. Ramanathan, J. Zao, Survivable Mobile WirelessNetworks: Issues, Challenges, and Research Directions,BBN Technologies, Cambridge, MA, 2002, September.

[10] J.S. Baras, H. Mehta, A probabilistic emergent routingalgorithm for mobile ad hoc networks, in: WiOpt’03:Modeling and Optimization in Mobile, Ad Hoc and WirelessNetworks, March 3–5, 2003.

[11] G.D. Caro, M. Dorigo, AntNet: A Mobile Agents Approachto Adaptive Routing, Universite Libre de Bruxelles, Bel-gium, Tech. Rep. IRIDIA/97-12, 1997.

[12] D. Camara, A.A.F. Loureiro, A GPS/ant-like routingalgorithm for ad hoc networks, IEEE Wireless Communica-tions and Networking Conference (WCNC?00), Chicago, IL(September) (2000).

[13] M. Heissenbttel, T. Braun, Ants-Based Routing in LargeScale Mobile Ad-Hoc Networks, University of Bern, 2003,Tech. Rep.

[14] M. Gunes, U. Sorges, I. Bouazizi, ARA—The ant-colonybased routing algorithm for MANETs, in: InternationalConference on Parallel Processing Workshops (ICPPW’02),Vancouver, BC, Canada, August 18–21, 2001.

[15] F. Ducatelle, G. Di Caro, L.M. Gambardella, Using antagents to combine reactive and proactive strategies forrouting in mobile ad-hoc networks, IDSIA/USI-SUPSI,Dalle Molle Institute for Artificial Intelligence Galleria 2,6928 Manno, Switzerland, Technical Report IDSIA 28-04,2005.

http://citeseer.nj.nec.com/perkins94highly.html

http://citeseer.nj.nec.com/perkins94highly.html


ARTICLE IN PRESS

[16] G. Di Caro, F. Ducatelle, L.M. Gambardella, AntHocNet:an ant-based hybrid routing algorithm for mobile ad hocnetworks, in: In Proceedings of the PPSN VIII—EighthInternational Conference on Parallel Problem Solving fromNatureLecture Notes in Computer Science, 3242, Springer-Verlag, Birmingham, UK, 2004, best paper award.

[17] H.F. Wedde, M. Farooq, T. Pannenbaecker, B. Vogel,BeeAdHoc: an energy efficient routing algorithm for mobilead hoc networks inspired by bee behavior, in: GECCO, June2005.

[18] M. Dorigo, G.D. Caro, L.M. Gambardella, Ant algorithmsfor discrete optimization, Universite Libre de Bruxelles,Tech. Rep. IRIDIA/98-10, 1999.

[19] S. Rajagopalan, C.-C. Shen, ANSI: A unicast routingprotocol for mobile ad hoc networks using swarm intelli-gence, in: International Conference on Artificial Intelligence(ICAI), Las Vegas, NV, USA, June 27–30, 2005.

[20] E.M. Royer, P.M. Melliar-Smith, L.E. Moser, An analysis ofoptimum node density for ad hoc mobile networks, in: IEEEInternational Conference on Communications, June 2001.

[21] W.R. Stevens, TCP/IP IllustratedThe Protocols, vol. 1,Addison Wesley Professional Computing Series, Reading,Massachusets, 1994.

[22] Z. Fu, P. Zerfos, H. Luo, S. Lu, L. Zhang, M. Gerla, Theimpact of multihop wireless channel on TCP throughput andloss, in: INFOCOM, April 2003.

Technology

A swarm intelligence-based unicast routing protocol for hybrid ad hoc networks