Comparative Performance Evaluation of Routing Protocols

for Mobile, Ad hoc Networks

Samir R. Das, Robert Castañeda and Jiangtao Yan Rimli Sengupta

Division of Computer Science Department of Computer Science
The University of Texas at San Antonio Rose-Hulman Institute of Technology
San Antonio TX 78249-0667 Terre Haute, IN 47803
U.S.A. U.S.A.

Abstract Agency, DARPA worked on PRNET (Packet radio Net-

work) [13] and SURAN (Survivable Adaptive Networks)
We evaluate several routing protocols for mobile, wire- [24] projects. They supported automatic route set up and
less, ad hoc networks via packet level simulations. The pro- maintenance in a packet radio network with moderate mo-
tocol suite includes routing protocols specifically designed bility. Interest in such networks has recently grown due to
for ad hoc routing, as well as more traditional protocols, the common availability of wireless communication devices
such as link state and distance vector, used for dynamic net- that can connect laptops and palmtops and operate in license
works. Performance is evaluated with respect to fraction of free radio frequency bands (such as the Industrial-Scientific-
packets delivered, end-to-end delay and routing load for a Military or ISM band in the U.S.). In an interest to run in-
given traffic and mobility model. It is observed that the new ternetworking protocols on ad hoc networks, a new working
generation of on-demand routing protocols use much lower group for Mobile, Ad hoc Networking (MANET) has been
routing load. However, the traditional link state and dis- formed within the Internet Engineering Task Force (IETF)
tance vector protocols provide, in general, better packet de- [17], whose charter includes developing a framework for
livery and delay performance. running IP based protocols in ad-hoc networks. Interest has
also been partly fueled by the recent IEEE standard 802.11
[5] that includes the MAC and physical layer specifications
for wireless LANs without any fixed infrastructure.
1. Introduction
Routing protocols in packet-switched networks tradition-
A mobile, ad hoc network [4] is an autonomous system ally use either link-state or distance-vector routing algo-
of mobile hosts connected by wireless links. There is no rithm [14]. Both algorithms allow a host to find the next
static infrastructure such as base stations. If two hosts are hop neighbor to reach the destination via the “shortest path.”
not within radio range, all message communication between The shortest path is usually in terms of the number of hops;
them must pass through one or more intermediate hosts that however, other suitable cost measures such as link utiliza-
double as routers. The hosts are free to move around ran- tion or queueing delay can also be used. Such shortest path
domly, thus changing the network topology dynamically. protocols have been successfully used in many dynamic
Thus routing protocols must be adaptive and able to main- packet switched networks. Prominent examples include use
tain routes in spite of the changing network connectivity. of link state protocol in OSPF (Open Shortest Path First)
Such networks are very useful in military and other tactical [18] and use of distance vector protocol in RIP (Routing In-
applications such as emergency rescue or exploration mis- formation Protocol) [11] for interior routing in the Internet.
sions, where cellular infrastructure is unavailable or unreli- Even though, any such protocol would, in principle, work
able. Commercial applications are also likely where there is for ad hoc networks, a number of protocols has been specif-
a need for ubiquitous communication services without the ically developed for use with ad hoc networks. The primary
presence or use of a fixed infrastructure. Examples include motivation is that the shortest path protocols, either link-
on-the-fly conferencing applications, networking intelligent state or distance vector, take too long to converge and have a
devices or sensors etc. high message complexity [4]. Because of the limited band-
Interest in such dynamic wireless networks is not new. It width of wireless links, message complexity must be kept
dates back to the seventies, when the U.S. Defense Research low. Also, potentially rapidly changing topology makes it
important to find routes quickly, even if the route may be the network topology and applies a shortest path algorithm
suboptimal [4]. (Dijkstra’s shortest path algorithm [6] in our simulations) to
Several new ad hoc routing protocols have been de- choose the next hop to a destination. Asynchronous link cost
veloped with this basic philosophy. They, however, vary updates may give rise to short-lived routing loops; however,
widely in characteristics. For example, some of these pro- they disappear by the time update messages have propagated
tocols are variations of distance vector routing. Some pro- throughout the network [14]. In our simulation we used the
tocols explicitly maintain redundant routing paths so that SPF implementation of link state protocol as described in
alternatives are available when a route changes. Some re- [25].
cently proposed protocols use a reactive approach for route
discovery and maintenance, instead of the more traditional, 2.2. Distance Vector Protocols
proactive approach [10]. In a reactive approach protocols
are “source initiated;” routes are discovered and maintained 
In the distance vector approach, for each destination ,
on an as needed basis, thus circumventing large overheads of every node  maintains a set of distances or costs,   ,
always maintaining routes between all possible source and 
where ranges over the neighbors of . Node is treated
destination pairs. The protocols are briefly reviewed in the 
as the next hop node for a data packet destined for , if
following section.    min    . To keep these distances up-to-
Even though many protocols have been proposed, their date, whenever there is any change of this minimum distance
comparative performance is not well understood. Current because of link cost changes, the new minimum distance is
literature reports only a limited amount of performance reported to the neighboring nodes. If, as a result, a minimum
study and when performance is reported, typically com- distance to any neighbor changes, this process is repeated.
parison has been made only to a selected few protocols This technique is the classical distributed Bellman-Ford al-
(typically only to link state and distance vector protocols). gorithm [2].
Specifically, the protocols proposed for mobile, ad hoc rout- Routing loops, both short-lived and long-lived, are pos-
ing have not been evaluated against one another. Our goal in sible in distributed Bellman-Ford. There is also a possibility
this paper to address this inadequacy by a thorough perfor- of counting-to-infinity problem, where it takes a large num-
mance study of several key protocols in the same framework ber of update messages to detect that a node is unreachable
to better understand their comparative merits and suitabil- [14]. Several protocols have been proposed to avoid long-
ity for deployment under different scenarios. To this end we lived loop and counting to infinity problems. They typically
use an existing, packet level, routing simulator called MaRS work by increasing the amount of information exchanged
[1] with added mobility modeling capability. A suite of ad between nodes or providing some sort of inter-nodal coordi-
hoc routing protocols is evaluated under varying mobility nation. For example, in the Border Gateway Protocol (BGP)
and traffic models. Traditional link state and distance vec- the entire path between the source and the destination is sent
tor protocols are included in the suite to provide a point of instead of just the distance [23]. In DUAL (Distributed Up-
comparison. date Algorithm) [8] inter-nodal coordination is achieved via
The rest of the paper is organized as follows. In Section a technique known as diffusing computation.
2, the dynamic routing protocols in packet networks are re- We focus our attention on two distance vector proto-
viewed with a special emphasis on the protocols evaluated in cols. The first, extended Bellman-Ford achieved good per-
this paper. Section 3 describes the simulation platform used formance for stationary networks in earlier simulation stud-
in evaluation and Section 4 presents the performance results. ies [25]. The second, DSDV [22], was specifically proposed
Conclusions are presented in Section 5. for mobile, ad hoc networks.

2. Routing Protocols for Ad hoc Networks Extended Bellman Ford Extended Bellman-Ford [3]
augments the classical Bellman-Ford by maintaining on
2.1. Link State Protocols node  , (in addition to the set of distances    ) a set of

nodes    , which immediately precede the destination

in the path from  to via neighbor . Then it is possible for
Each node maintains its own view of the network topol- 
ogy, including link costs of all its outgoing links. To keep the source node to construct the whole path to the destina-
views up-to-date, each node broadcasts the link costs of all tion, by repeatedly using the preceding node   as a new
its neighbors to all other nodes in the network using flood- destination. It can be shown that the protocol is free from
ing. This is done whenever there is a change in link costs. both long-lived loop and counting-to-infinity problems, if
As a node receives this information, it updates its view of each node avoids sending route change updates to a neigh-
bor for any destination  , if that neighbor is in the path to


In a mobile, ad hoc network any node can potentially be a neighbor.  .
 We used the EXBF implementation described in [25] for comes unreachable because of a network partition, the pro-
our evaluation. Here, periodically or whenever a failure or tocol erases (route erasure phase) all invalid routes.
reconnect occurs, link costs are recalculated, and if there is a TORA uses a notion of node “height” to maintain the des-
change in the minimum distance, the new minimum distance tination oriented DAG. Each node maintains a height and
is reported to the neighboring nodes. Several protocols pro- exchanges this value with each neighbor. The significance
posed in the literature are based on a similar idea of main- of the height is that a link is always directed from a “higher”
taining the second-to-last hop (predecessor) for the shortest node to a “lower” node. Note that this notion of height and
path to each destination to achieve loop freedom. See [19] link directions are destination specific. Independent copies
and the references therein. of the protocol runs for each possible destination node in the
network.
DSDV The destination sequenced dis- In the initial route construction phase, the height of a
tance vector or DSDV protocol [22] has been specifically node carries the notion of distance (in hops) of the node from
targeted for mobile networks. DSDV augments the classi- the destination. However, this distance information is even-
cal, distributed Bellman-Ford by tagging each distance en- tually lost during route maintenance phase. Since multiple
try    by a sequence number that originated in the des- routes are maintained in TORA, an obvious question is the
 choice of route. Two alternatives are suggested – choosing a
tination node . Each node maintains this sequence number,
incrementing it each time the node sends an update to the neighbor randomly so that the loads are more or less evenly
neighbors. The sequence number is disseminated in the net- distributed or choosing the lowest neighbor [20]. We have
work via update messages. The destination sequence num- chosen the latter in our simulations.
ber is used to determine the “freshness” of a route. Always
the latest sequence number is used for updating routes. For 2.4. On Demand Protocols
equal sequence numbers, the one with the smallest distance
metric is used. It has been shown that DSDV avoids long- Link state and distance vector protocols are primarily
lived loops and counting to-infinity problems [22]. How- proactive protocols in the sense that routes are maintained
ever, in our knowledge its performance has not yet been re- to all potential destinations (possibly all nodes in the net-
ported in literature. work) all the time, whether are not all such routes are actu-
ally used [10]. Route maintenance can be a large overhead
2.3. Multipath Protocols – TORA because of a significant amount of route update traffic, es-
pecially for large networks. Reactive or on-demand proto-
The unique feature of the temporally ordered routing al- cols, on the other hand, create and maintain routes only on
gorithm or TORA [20] is maintaining multiple routes to the “as needed” basis. Thus, when a route is needed, some sort
destination so that many topological changes do not require of global search procedure is employed. The family of clas-
any reaction at all, as having just a single route is sufficient. sical flooding algorithms belong to the reactive group. Two
The protocol reacts only when all routes to the destination examples of recently proposed reactive protocols are DSR
are lost. In that case routes are re-established via a tempo- and AODV. Note that TORA, described earlier is also par-
rally ordered sequence of diffusing computations, which are tially reactive in the sense that route creation is initiated on
essentially link reversals (to be described momentarily). In demand. However, route maintenance is done on a proac-
the event of network partitions, the protocol is able to detect tive basis such that multiple routing options are available in
the partition and erase all invalid routes. case of link failures.
TORA is based, in part, on the classical work by Gafni
and Bertsekas [7], who consider a similar problem of main- DSR Dynamic source routing or DSR [12] uses a tech-
taining a destination oriented directed acyclic graph (DAG) nique where the source of a data packet determines the
in the face of topological changes. A DAG is considered complete sequence of nodes through which to forward the
destination oriented, if for every node, there is a path to packet; the source explicitly lists this route in the packet’s
a given destination. If link failures make such a graph header. DSR builds routes on demand using flooded query
“destination disoriented,” a series of link reversals ensue packets that carry the sequence of hops they passed through.
so that the graph again becomes destination oriented in fi- Once a query reaches the destination, destination replies
nite time. The graph is initially constructed (route discovery with a reply packet that simply copies the route from the
or construction phase) in a “source-initiated” fashion, us- query packet and traverses it backwards. Each node has
ing a query flood followed by update routing packets. From

We assume that the wireless links are symmetric, which may not be the
that point it is maintained (route maintenance phase) using
case in practice. DSR can tolerate asymmetric links by, for example, using
link reversals alone, whenever topological change causes a an independentroute discovery from destination back to the source. We did
node to lose its last downstream link. If the destination be- not consider asymmetric links in our evaluations in this work.
a route cache, where complete routes to desired destina- quenching of route requests is not used) and the source node
tions are stored as gleaned from the response packets. These (and not an intermediate node) is always the initiator of the
routes are used for data packets. Route failure is detected route request, the initial route may the shortest. But depend-
by the failure of an attempted message transmission. Such a ing on the changes in topology this route may not always re-
failure initiates an error packet sent backward to the source. main the shortest.
The error packet erases all routes in the route caches of all
intermediate nodes on its path, if the route contains the failed 2.5. Other Protocols
link.
DSR has an unique advantage by virtue of source rout- Several other protocols have appeared in literature for
ing. As the route is part of the packet itself, routing loops, mobile, ad hoc networks. Zone routing protocol (ZRP) [10]
either short- or long-lived, cannot be formed as they can be is a zone or cluster based routing protocol that is a hybrid be-
immediately detected and eliminated. This property opens tween proactive and reactive routing. It is targeted for very
up the protocol to a variety of useful optimizations. For ex- large networks and divides the network into zones or clusters
ample, a flooded query can be quenched early by having any of nodes. The nodes within a zone is close to one another.
non-destinationhost reply to the query if that host has a route Use of proactive routing is advocated within a zone and
to the intended destination. Also, routes can be improved reactive routing across zones [10]. Cluster based routing,
by having nodes promiscuously listen to conversations be- however, is not new. Quite a few cluster-based approaches
tween other nodes in proximity. have appeared in the past. See, for example, [24, 9, 16].
We have not yet included ZRP in our suite of protocols
AODV Ad hoc, on-demand distance vector protocol or considered for evaluation because of a couple of reasons.
AODV [21] is an on-demand variation of distance vector ZRP can be viewed more a “routing framework” rather than
protocols. AODV uses destination sequence numbers like an independent protocol, as potentially any proactive proto-
DSDV to determine freshness of routing information. In col can be employed for intrazone routing and any reactive
AODV, flooded requests are used to create route, with the protocol can be employed for interzone routing. Also, ZRP
destination responding to the first such request, much as is suitable for a very large network such as a good number
in DSR. However, AODV maintains routes in a distributed of clusters can be formed, while our study so far is concen-
fashion, as routing table entries, on all intermediate nodes trated on moderately sized network. Note that TORA, DSR,
on the route. Routing table entries are tuples in the form AODV and ZRP are the four protocols currently under study
of destination, next hop, distance . Nodes forwarding


by the IETF MANET working group as candidate protocols
queries remember the earlier hop taken by the query packet. for evaluation and standardization [17].
This hop is used to forward the reply packet back to the
source. The reply packet sets up the routing table entries on 3. Simulation Model
its path. AODV advocates use of “early quenching” of re-
quest packets, i.e., any node having a route to the destination A discrete event, packet-level, routing simulator called
can reply to a request. AODV also uses a technique called MaRS (Maryland Routing Simulator) [1] was used for com-
route expiry, where a routing table entry expires after a pre- parative performance evaluation. MaRS is a flexible plat-
determined period, after which fresh route discovery must form developed specifically for evaluation and compari-
be initiated. son of network routing algorithms. MaRS was used previ-
AODV maintains the addresses of the neighbors through ously for comparative evaluation of link-state and distance-
which packets destined for a given destination were re- vector routing protocols for the NSFNET T1 backbone net-
ceived. A neighbor is considered active (for a destination), work with the possibility of link failures [25]. We aug-
if it originates or relays at least one packet for that destina- mented MaRS to provide node mobility. The nodes can
tion, within the past active timeout period. A routing table move around in a rectangular region according to a given
entry is active if it is used by an active neighbor. The path mobility model (to be described momentarily). Each node
from a source to a destination via the active routing table en- has a fixed radio range and has a link to every other node in
tries is called an active path. On a link failure, all routing the system. If the other node is not within range the link cost
table entries are erased for which the failed link is on the is infinity. Otherwise, the link cost is modeled by the hop-
active path. This is accomplished by an error packet going normalized delay function, same as the revised ARPAnet
backwards to the active neighbors, which forward them to cost metric [15, 25].
their active neighbors and so on. This technique effectively Each node is modeled by a store-and-forward, queueing
erases the route backwards from the failed link. station, and is characterized by parameters such as buffer
Neither DSR nor AODV guarantees shortest path. If the space and processing speed. Each link is characterized by
destination alone can respond to route requests (i.e., early a bandwidth and propagation delay. A link is modeled as an
FCFS queue with service time as the transmission time. Cur- always unicast. Routing packets can be broadcast or uni-
rently, our study is limited to network layer details. Thus, no cast depending on the protocol requirement.
link layer details, such as MAC protocol, multiple-access in- All nodes are assumed to have adequate buffer capacity
terference or link errors, are modeled, nor are any physical, for buffering packets awaiting forwarding. Data packets are
radio channel level details. processed (includes parsing the header, consulting the rout-
The routing protocol is modeled as an independent rout- ing table or cache and adding the packet to the appropriate
ing module, one at each node, which maintains routing in- outgoing packet queue) in parallel. Data packet processing
formation (such as next-hops, distances, routing table etc. costs are fixed (1 ms). Routing packets have higher prior-
depending on the protocol used) and responds to routing ity over data packets in the node’s outgoing packet queue.
packets and link status changes. Routing packets are distinct Routing packets are processed sequentially. Routing packet
from data packets in the simulator and are used for route processing cost and routing packet sizes depend on the rout-
maintenance. The nodes forward data packets via the next ing protocol being used. Data packet sizes are defined by
hop link as per the routing information provided by the rout- the workload model plus a fixed, small header. However,
ing module. If the next hop link is broken or there is no for source routing the header length is variable and can be
next hop information available, data packets are dropped un- long depending on the length of the route.
til some usable next hop information is available. In source
routed protocols, however, the data packets themselves con- Mobility Nodes move around in a rectangular region of
tain the route gleaned from the route cache maintained by size 1000 m
1000 m according to a mobility model. The
the routing module. nodes have a constant radio range of 350 m. Nodes are
Workload is defined in terms of connections. A con- constantly moving, thus putting stress on the routing pro-
nection is a unicast conversation between a source and a tocols. The node movements, however, are discretized for
sink. The source and sink are modules associated with ease of modeling in a discrete event framework. Each node
nodes. Several workload models are provided in MaRS. chooses a direction, speed and distance of move based on a
In this paper, however, we use the simplest model, where pre-defined distribution and then computes its next position

the source generates data packets destined for the sink at a and the time instant  of reaching that position. Simi-
steady rate. This traffic is characterized by a packet length larly, a new “move” is again computed at simulation time
and a random (exponentially distributed in our simulation)  . A node computes its neighborhood after each such move,
inter-packet generation interval. There is no flow or conges- thus generating link failure and link repair events that in turn
tion control. drive the routing protocol.
For the experiments described in this paper, the speed of
3.1. Detecting link status changes each move is uniformly distributed between a given range
(0.4 – 0.6 m/sec, low mobility experiments and 3.5 – 4.5 m
An important feature of mobile networks is detection of /sec, high mobility experiments), distance is exponentially
link failures or appearances. This can be done in a few ways, distributed with a mean of 5 m, and the direction is uni-
such as periodic link status sensing/probing by so-called formly distributed within    with respect to the
hello messages. Link layer protocols that use acknowledg- direction of the previous move.
ments can also be used to detect link failures. Sine no link
layer details are modeled, a link layer event is generated au- Workload A simple workload model is used. All data
tomatically whenever a link fails or reappears, i.e., a node packets are 512 bytes long, and interarrival times are expo-
goes out or in range. The routing protocol responds to this nentially distributed with a mean of 300 ms. There is no ac-
event. No hello messages have been modeled. knowledgment or flow or congestion control in the work-
load model. Flow or congestion control mechanisms will
3.2. Simulation Parameters be influenced by the routing dynamics and thus will change
the load on the network. It is not clear how this will influ-
Physical network We assume a channel bandwidth of 1.5 ence our performance metrics. Workload traffic is always
Mbits/sec. Since no multiple-access contention or interfer- between a pair of source and sink nodes, called a connec-
ence is modeled, each link essentially enjoys the entire chan- tion. The number of such pairs or connections is varied over
nel bandwidth while transmitting packets. In the simulation a wide range in the simulation experiments. In the perfor-
model, a packet can be unicast (received only by a specific mance plots, it is presented in terms of no. of connections
neighbor) or broadcast (received by all neighbors). Broad- per node in the network.
cast transmissions are modeled as a sequence of unicast 
Passive eavesdroppingmay improve performance of some routing pro-
transmissions on all active links of a node, but the packet is tocols such as DSR. This could be modeled using broadcast transmission.
counted only once in simulation statistics. Data packets are However, we did not model eavesdropping yet.
 1 1
0.95 0.95
fraction of packets delivered

fraction of packets delivered

0.9 0.9
0.85 0.85
0.8 0.8
0.75 0.75
0.7 SPF 0.7 SPF
EXBF EXBF
0.65 DSDV 0.65 DSDV
TORA TORA
0.6 DSR 0.6 DSR
0.55 AODV 0.55 AODV

0.5 0.5
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
#connections/node #connections/node

Figure 1. Fraction of packets delivered for all Figure 2. Fraction of packets delivered for all
routing protocols for the low mobility case. routing protocols for the high mobility case.

4. Performance Results is assigned from the current node’s view of the state of all
links in the network. In EXBF and DSDV, an alternate min-
We have simulated a 30 node mobile, ad hoc network imum cost route is found via a different neighbor. However,
with respect to the above mobility and workload models. All no such alternative is available for DSR and AODV and thus
protocols are studied with respect to three key performance packets are dropped until route can be repaired. TORA, sur-
metrics: prisingly, offers the lowest packet delivery fraction in spite
Fraction of packets delivered: measured as a ratio of of its multipath capability. In our observation, the key rea-
the number of data packets delivered to the destination son for this is that the initial route discoveries take longer.
and the number of data packets sent by the sender. This affects the performance most when there is a reconnect
after a network partition. Also, TORA is very sensitive to
End-to-end delay: measured in ms. the loss of routing packets compared to the other protocols.
This more than offsets the advantages gained by the multi-
Routing load: measured in a normalized fashion in
path nature of the protocol.
terms of number of bytes of routing packets transmitted
We note here that buffering data packets while route dis-
per byte of data packets transmitted. The latter includes
covery in progress has a great potential to improve DSR,
only the data packets finally delivered at the destina-
AODV and TORA performances. However, this alternative
tion and not the ones that are dropped. The transmis-
has not been evaluated. We also have not used early quench-
sion on each hop is counted once for both routing and
ing of route request packets by a non-destination node in
data packets. This gives an idea of network bandwidth
DSR and AODV. We have noticed that AODV performs
consumed by routing packets with respect to “useful”
very poorly by picking up stale routes, if early quenching
data packets.
is used. It affects both its packet delivery and delay perfor-
The first set of figures present the fraction of packets de- mance significantly. Thus we feel that having the destina-
livered for all protocols for low (Figure 1) and high (Fig- tion always reply to route requests provides a fairer com-
ure 2) mobilities. Note the excellent behavior on the part of parison. AODV has a slightly worse packet delivery perfor-
all link state and distance vector protocols, but considerably mance than DSR because of higher drop rates. AODV uses
lower packet delivery fraction for on-demand and multipath route expiry, dropping some packets when a route expires
protocols. On-demand protocols (DSR and AODV) drop a and a new route must be found.
considerable number of packets during the route discovery The average end-to-end delays are shown in Figures 3
phase, as route acquisition takes time proportional to the dis- and 4 for low and high mobility, respectively. The short-
tance between the source and destination. The situation is est path protocols (SPF, EXBF and DSDV) show the min-
similar with TORA. Packet drops are fewer with proactive imum delay characteristics. AODV and DSR show worse
protocols as alternate routing table entries can always be as- characteristics as their routes are typically not the shortest.
signed in response to link failures. In SPF, an alternate route Even if the initial route discovery phase finds the shortest
 10
9
8
7 10

6 9
delay (ms)

5 SPF 8
EXBF
4 DSDV 7
TORA
6

delay (ms)
3 DSR
AODV
2 5 SPF
EXBF
1 4 DSDV
TORA
0 3 DSR
0 1 2 3 4 5 6 7 8 9 10 AODV
#connections/node 2
1
Figure 3. Average end-to-end delay for all
0
routing protocols for the low mobility case. 0 1 2 3 4 5 6 7 8 9 10
#connections/node

Figure 4. Average end-to-end delay for all

route (it typically will), the route may not remain the shortest routing protocols for the high mobility case.
over a period of time due to node mobility. Also, note that
in AODV routes are maintained as a soft state, i.e., routes
expire after a timeout interval and fresh route discovery is
initiated. Accordingly, AODV performs a little better delay-
wise and can possibly do even better with some fine-tuning
of this timeout period by making it, for example, a function
of node mobility. TORA has the worst delay characteristics
because of the loss of distance information with progress.
The routing load characteristics shown in Figures 5 and
6 are interesting. Note that the routing load varies over a 40
very wide range and hence the plots use a logarithmic scale 20 SPF
EXBF
for the vertical axis. SPF expends significantly more rout- 10 DSDV
7 TORA
ing load than the other protocols. The distance vector pro- 5
routing load (normalized)

DSR
tocols, EXBF and DSDV, have very similar routing loads, 3 AODV
2
and much lower than SPF. DSR and AODV perform very
well, particularly for smaller number of connections. TORA 1
0.7
again performs much worse than expected for the similar 0.5
0.3
reasons mentioned earlier in connection with packet deliv- 0.2
ery fraction. 0.1
It appears that the theoretical “worst” case communica- 0.07
0.05
tion complexity (number of messages required to adapt to 0.03
a link failure/recovery) does not provide much insight into 0.02
the average case behavior obtained via simulation. For ex- 0.01
0 1 2 3 4 5 6 7 8 9 10
ample, SPF has a worst case communication complexity #connections/node
of 
 [14], where  is the number of links in the net-
work. On the other hand communication complexity in pro- Figure 5. Normalized routing load for all rout-
tocols based on distributed Bellman-Ford, such as EXBF ing protocols for the low mobility case.
and DSDV, is exponential in the number of nodes  in the
network [3], which should be much higher. However, SPF
has a much higher routing load in the simulations.
Most protocols benefit to some degree as the number of
connections grow large. This is because a single route re-
 40 higher routing load. The on-demand protocols suffer from
20 SPF
EXBF sub-optimal routes as well as worse packet delivery fraction
10 DSDV because of more dropped packets. However, they are sig-
7 TORA
5 nificantly more efficient in terms of the routing load. The
routing load (normalized)

DSR
3 AODV multipath protocol, TORA, did not perform well in spite of
2
maintaining multiple redundant path. The overhead of find-
1
0.7 ing and maintaining multiple paths and the protocol’s sen-
0.5
sitivity to the loss of routing packets seem to outweigh the
0.3
0.2 benefits of multiple paths. Also, the end-to-end delay per-
0.1 formance is poor because of the loss of distance information.
0.07 The routing load differentials between all routing protocols
0.05
0.03 reduce with large number of peer-to-peer conversations in
0.02 the network. However, the other performance differentials
0.01 are not affected conclusively.
0 1 2 3 4 5 6 7 8 9 10
#connections/node It is important to note the limitations of the study. First,
a packet level simulation has its own limitations. No mul-
Figure 6. Normalized routing load for all rout- tiple access interference is modeled. Thus a high routing
ing protocols for the high mobility case. packet load does not interfere as much with the data trans-
missions (except for queueing delays) as it would in real-
ity. The current study best reflects the performance when
pair can potentially benefit many connections. Thus rout- all active links in the network are on a separate frequency
ing load do not increase as much as the data load with in- band. Second, only a moderate size network has been stud-
creasing number of connections. Also note that the rout- ied. Though it is unclear what sizes will be realistic for
ing loads in DSDV, EXBF, DSR and AODV are in the same an ad hoc network running IP based protocols, using a few
ballpark as the number of connections becomes large. Thus other sizes, going upto a few hundreds, will provide more
the proactive distance vector protocols should be favored at maturity to the study. Third, a few different traffic models,
large number of connections as they provide better end-to- for example, dynamically changing peers for conversations
end delay characteristics. We also note in the passing, that and introduction of hot-spots should be studied to evaluate
DSR uses somewhat more bandwidth (10–20% in our exper- the sensitivity to traffic models. Fourth, fine tuning of cer-
iments) because of source routing that increases the size of tain protocol parameters (e.g., various timeout periods for
the header in data packets. Even counting this in as a part the on-demand protocols) is possible with changing mobil-
of routing load DSR is very competitive with AODV. How- ity and traffic characteristics. We have used reasonable val-
ever, this bandwidth usage is expected to increase for larger ues that work well, but have not changed the values for dif-
networks and may make DSR less attractive. ferent traffic and mobility. Also, certain protocol specific
optimizations (e.g., passive eavesdropping in DSR) as well
as more general optimizations (e.g., buffering of data pack-
5. Conclusions ets on route loss until route is repaired) are possible. They
may impact relative performance. Fifth, impact of memory
Our work is the first attempt towards a comprehensive usage by the protocols have been ignored. This may be im-
performance evaluation of routing protocols for mobile, ad portant as the computing nodes deployed in a mobile, ad hoc
hoc networks. We evaluated all but one protocol currently environment can be low power and small size devices. In
considered in the IETF MANET working group, in addition spite of these limitations, we have gained valuable insight
to more traditional link state and distance vector protocols. into the behavior of routing algorithms in an ad hoc network.
Steady state performance in terms or fraction of packets de- Our future work will address these limitations.
livered, delay and routing load have been considered as the
performance metrics.
Even with a packet-level simulation model the essential Acknowledgments
aspects of the routing protocols are exposed. The key ob-
servations are as follows. The proactive, shortest path pro-
tocols provide excellent performance in terms of end-to-end This work is partially supported by AFOSR grant no.
delays and packet delivery fraction, however, at the cost of F49260-96-1-0472, Texas Advanced Technology Program

As mentioned earlier the Zone Routing Protocol (ZRP) [10] has not grant no. 010115-248b and NSF CAREER award no. ASC-
been evaluated. 9733836.
References [20] V. D. Park and M. S. Corson. A highly adaptive distributed
routing algorithm for mobile wireless networks. In Proceed-
[1] C. Alaettinoglu, A. U. Shankar, K. Dussa-Zieger, and ings of IEEE INFOCOM’97 Conf., April 1997.
I. Matta. Design and implementation of MaRS: A routing [21] C. Perkins. Ad hoc on demand distance vector (AODV) rout-
testbed. Journal of Internetworking: Research and Experi- ing. http://www.ietf.org/internet-drafts/
ence, 5(1):17–41, 1994. draft-ietf-manet-aodv-00.txt, 1997. IETF In-
[2] D. Bertsekas and R. Gallager. Data Networks. Prentice-Hall, ternet Draft.
[22] C. E. Perkins and P. Bhagwat. Highly dynamic destination-
1987.
sequenced distance-vector routing (DSDV) for mobile com-
[3] C. Cheng, R. Riley, and S. P. R. Kumar. A loop-free extended
puters. In Proceedings of the ACM SIGCOMM ’94 Confer-
bellman-ford routing protocol without bouncing effect. In
ence, pages 234–244, August 1994.
Proc. of ACM SIGCOMM Conf., pages 224–236, 1989.
[23] Y. Rekhter and T. Li. A border gateway protocol-4. RFC
[4] M. Corson, S. Batsell, and J. Macker. Architectural consid-
827, March 1995.
erations for mobile mesh networking.
[24] N. Schacham and J. Westcott. Future directions in packet ra-
http://tonnant.itd.nrl.navy.mil/mmnet/
dio architectures and protocols. Proceedings of the IEEE,
mmnetRFC.txt, May 1996. Request for Comments Draft.
75(1):83–99, Jan 1987.
[5] I. S. Department. IEEE 802.11 draft standard for wireless
[25] A. U. Shankar, C. Alaettinoglu, K. Dussa-Zieger, and
LAN, MAC and physical layer specifications, May 1997.
I. Matta. Transient and steady-state performance of rout-
[6] E. W. Dijkstra. A note on two problems in connection with
ing protocols: Distance-vector versus link-state. Journal
graphs. Numerical Mathematics, 1:269–271, Oct. 1959.
of Internetworking: Research and Experience, pages 59–87,
[7] E. Gafni and D. Bertsekas. Distributed algorithms for gen-
1995. (Preliminary version appeared in Proc. ACM SIG-
erating loop-free routes in networks with frequently chang-
METRICS/PERFORMANCE Conf., 1992, pages 181-192.).
ing topology. IEEE Trans. on Communication, 29(1):11–18,
January 1981.
[8] J. J. Garcia-Luna-Aceves. A unified approach to loop free
routing using distance vectors or link states. In Proc. of ACM
SIGCOMM’89 Conf., pages 212–223, Sept. 1989.
[9] M. Gerla and J.-C. Tsai. Multicluster, mobile, multimedia
radio networks. Wireless Networks, 1, 1995.
[10] Z. J. Haas and M. R. Pearlman. The zone routing
protocol (ZRP) for ad hoc networks. http://www.
ietf.org/internet-drafts/draft-ietf-
manet-zone-zrp-00.txt, 1997. IETF Internet Draft.
[11] C. Hedrick. Routing information protocol. RFC 1058, June
1988.
[12] D. Johnson and D. Maltz. Dynamic source routing in ad hoc
wireless networks. In T. Imielinski and H. Korth, editors,
Mobile computing. Kluwer Academic, 1996.
[13] J. Jubin and J. D. Tornow. The DARPA packet radio network
protocols. Proceedings of the IEEE, 75(1):21–32, January
1987.
[14] S. Keshav. An Engineering Approach to Computer Network-
ing : ATM Networks, the Internet, and the Telephone Net-
work, chapter 11. Addison-Wesley, 1997.
[15] A. Khanna and J. Zinky. A revised ARPANET routing met-
ric. In Proc. of the ACM SIGCOMM’89 Conf., pages 45–56,
Sept. 1989.
[16] P. Krishna, N. H. Vaidya, M. Chatterjee, and D. K. Pradhan.
A cluster-based approach for routing in dynamic networks.
ACM SIGCOMM Computer Communication Review, April
1997.
[17] J. Macker and S. Corson.
Mobile ad hoc networks (MANET). http://www.ietf.
org/html.charters/manet-charter.html,
1997. IETF Working Group Charter.
[18] J. Moy. OSPF version 2. RFC 1247, July 1991.
[19] S. Murthy and J. J. Garcia-Luna-Aceves. An efficient routing
protocol for wireless networks. Mobile Networks and Appli-
cations, 1(2):183–197, Oct. 1996.