On-demand Multipath Distance Vector Routing in Ad Hoc Networks

Mahesh K. Marina Samir R. Das

Department of Electrical & Computer Engineering and Computer Science
University of Cincinnati
Cincinnati OH 45221-0030
U.S.A.
E-mail: marina, sdas@ececs .uc . edu

Abstract wide flood) is initiated whenever a route is needed. Sev-

eral performance studies [ 3 , 7 , 131 of ad hoc networks have
We develop an on-demand, multipath distance vector shown that on-demand protocols incur lower routing over-
protocol for mobile ad hoc networks. Speci$cally, we pro- heads compared to their proactive counterparts. However,
pose multipath extensions to a well-studied single path rout- they are not without performance problems. High route
ing protocol known as Ad hoc On-demand Distance Vector discovery latency together with frequent route discovery
(AODV). The resulting protocol is referred to as Ad hoc On- attempts in dynamic networks can affect the performance
demand Multipath Distance Vector (AOMDV). The protocol adversely. Multipath on-demand protocols try to allevi-
computes multiple loop-free and link-disjoint paths. Loop- ate these problems by computing multiple paths in a single
freedom is guaranteed by using a notion of “advertised route discovery attempt. Multiple paths could be formed at
hopcount.” Link-disjointness of multiple paths is achieved both traffic sources as well as at intermediate nodes. New
by using a particular property of Jlooding. Performance route discovery is needed only when all paths fail. This re-
comparison of AOMDV with AODV using ns-2 simulations duces both route discovery latency and routing overheads.
shows that AOMDV is able to achieve a remarkable im- Multiple paths can also be used to balance load by forward-
provement in the end-to-end delay - often more than a fac- ing data packets on multiple paths at the same time, though
tor of two, and is also able to reduce routing overheads by we will not explore this aspect in our work here.
about 20%.
Our goal here is to develop an on-demand, multipath
distance vector protocol as an extension to a well-studied
1 Introduction single path routing protocol known as Ad hoc On-demand
Distance Vector (AODV) [27, 281. We refer to the new
Mobile ad hoc networks are characterized by dynamic protocol as Ad hoc On-demand Multipath Distance Vector
topology due to node mobility, limited channel bandwidth (AOMDV) protocol. Primary design goal behind AOMDV
and limited battery power of nodes. The key challenge here is to provide efficient fault tolerance in the sense of faster
is to be able to route with low overheads even in dynamic and efficient recovery from route failures in dynamic net-
conditions. Overhead here is defined in terms of the rout- works. To achieve this goal, AOMDV computes “multiple
ing protocol control messages which consume both channel loop-free and link-disjoint paths.” The notion of an “ad-
bandwidth as well as the battery power of nodes for com- vertised hopcount” is used to maintain multiple loop-free
munication/processing. paths. A particular property of flooding is used to ensure
In order to reduce routing overheads, on-demand routing link-disjointness of the multiple paths computed within a
protocols build and maintain only needed routes. Examples single route discovery.
include Dynamic Source Routing (DSR) [ 141, Temporally
Ordered Routing Algorithm (TORA) [24] and Ad hoc On- The rest of the paper is organized as follows. In section
demand Distance Vector routing (AODV) [27, 281. This 2, we review the AODV protocol. In section 3 , we develop
is in contrast to proactive protocols (e.g., Destination Se- the AOMDV protocol as an extension of AODV. In section
quenced Distance Vector (DSDV) [26]) that maintain routes 4, we evaluate the performance of AOMDV using ns-2 sim-
between all node pairs all the time. In on-demand proto- ulations. Related work is presented in section 5, and con-
cols, a route discovery process (typically via a network- clusions in section 6.

1092-1658/01 $17.00 0 2001 IEEE 14

2 Ad Hoc On-Demand Distance Vector Rout- if ( seqnumf < s e q n u m f ) or (1)
ing ( ( s e q n u m t = seqnum!) and
(hopcount: > hopcount?)) then
AODV [27,28] combines the use of destination sequence
numbers in DSDV [26] with the on-demand route discovery seqnumf := s e q n u m f ; (2)
technique in DSR [ 141 to formulate a loop-free, on-demand, hopcount! := hopcounty + 1; (3)
single path, distance vector protocol. Unlike DSR, which nexthop! := j; (4)
uses source routing, AODV is based on hop-by-hop rout-
ing approach. Below we review some of the key features of endif
AODV to provide sufficient background for AOMDV de-
scribed in the next section. Figure 1. AODV route update rule. This is used
whenever a node i receives a route advertisement to
2.1 Route Discovery and Route Maintenance a destination d from a neighbor j. The variables
seqnum:, hopcount: and nexthopf represent the
When a traffic source needs a route to a destination, it ini- destination sequence number, hopcount and the nex-
tiates a route discovery process. Route discovery typically thop, respectively, for a destination d at node i.
involves a network-wide flood of route request (RREQ)
packets targeting the destination and waiting for a route
reply (RREP). An intermediate node receiving a RREQ
two different nodes for the same destination. The AODV
packet first sets up a reverse path to the source using the
protocol maintains an invariant that destination sequence
previous hop of the RREQ as the next hop on the reverse
numbers monotonically increase along a valid route, thus
path. If a valid route to the destination is available, then the
preventing routing loops. This is explained below further,
intermediate node generates a RREP, else the RREQ is re-
as it will play a crucial role in the understanding of the mul-
broadcast. Duplicate copies of the RREQ packet received
tipath protocol.
at any node are discarded. When the destination receives a
A node can receive a routing update via a RREQ or
RREQ, it also generates a RREP. The RREP is routed back
RREP packet either forming or updating a reverse or for-
to the source via the reverse path. As the RREP proceeds to-
wards the source, a forward path to the destination is estab- ward path. In the rest of the paper, we refer to such routing
updates received via a RREQ or RREP as “route advertise-
lished. Several optimizations have been proposed in the lit-
ments.’’ The update rule in Figure 1 is invoked upon re-
erature to contain the scope of the flood [5] and to reduce the
ceiving a route advertisement. It is easy to see why loops
redundancy of the broadcasts during the flood [ 2 2 ] . Since
cannot be formed if this rule is followed. Consider the tu-
these are somewhat orthogonal to our interest, we will limit
ple (-seqnum:, hopcount:) where seqnum! represents
our discussion to only pure flooding in this paper. Further,
we assume bidirectional links and a connected network. the sequence number at node i for the destination d. Simi-
larly, hopcount: represents the hopcount to the destination
Route maintenance is done using route error (RERR)
d from node i . For any two successive nodes i and j on a
packets. When a link failure is detected (by a link layer
feedback, for example), a RERR is sent back via separately valid path to the destination, j being the next hop from i to
d, the route update rule in Figure 1 enforces that
maintained predecessor links to all sources using that failed
link. Routes are erased by the RERR along its way. When ( - s e q n u m t , hopcount:) > (-seqnum;, hopcount;),
a traffic source receives a RERR, it initiates a new route
discovery if the route is still needed. Unused routes in the where the comparison is in the lexicographic sense. Thus,
routing table are expired using a timer-based technique. the tuples (-seqnum:, hopcount:) along any valid route
are in a lexicographic total order, which in turn implies loop
2.2 Sequence Numbers and Loop Freedom freedom.

Sequence numbers in AODV play a key role in ensur- 3 Ad Hoc On-Demand Multipath Distance
ing loop freedom. Every node maintains a monotonically Vector Routing
increasing sequence number for itself. It also maintains the
highest known sequence numbers for each destination in the The main idea in AOMDV is to compute multiple paths
routing table (called “destination sequence numbers”). Des- during route discovery. It is designed primarily for highly
tination sequence numbers are tagged on all routing mes- dynamic ad hoc networks where link failures and route
sages, thus providing a mechanism to determine the relative breaks occur frequently. When single path on-demand rout-
freshness of two pieces of routing information generated by ing protocol such as AODV is used in such networks, a new

15
route discovery is needed in response to every route break.
Each route discovery is associated with high-overhead and
latency. This inefficiency can be avoided by having multi-
ple redundant paths available. Now, a new route discovery
I destination
sequence number
I
I
destination
seauence number
advertisedhopcount
I 1
is needed only when all paths to the destination break. routelist
A noteworthy feature of the AOMDV protocol is the use { (nezthopl, hopcountl),
of routing information already available in the underlying I expiration-timeout I
AODV protocol as much as possible. Thus little additional
overhead is required for the computation of multiple paths. (a) AODV
The AOMDV protocol has two main components: (b) AOMDV

1. a route update rule to establish and maintain multiple

loop-free paths at each node. Figure 2. Structure of routing table entries for AODV
and AOMDV.
2. a distributed protocol to find link-disjoint paths.
In the following, we describe each of these two components
in detail.
Figure 2 shows the structure of the routing ta-
ble entries for AODV and AOMDV. In AOMDV,
3.1 Computing Multiple Loop-free Paths advertised-hopcount replaces hopcount in AODV. A
route-list replaces the nexthop, and essentially defines
Observe that each route advertisement arriving at a node
multiple next hops with respective hopcounts. However, all
during AODV route discovery potentially defines an alter-
next hops still have the same destination sequence number.
nate path to the source or the destination. For example, each
The advertised-hopcount is initialized each time the
copy of the RREQ packet arriving at a node defines an al-
sequence number is updated. A node i updates its
ternate path back to the source. However, accepting all such
advertised-hopcount for a destination d whenever it sends
copies naively to construct routes will lead to routing loops.
a route advertisement for d. Specifically, it is updated as fol-
To see how loops can occur, consider this simple example.
lows:
Source S initiates a flood of RREQ packets. An interme-
diate node A broadcasts the RREQ. A neighbor B rebroad- advertised-hopcount! := maxk { hopcountk I
casts it, which in turn is heard by A. If A accepts this RREQ
copy to form a reverse path, this will form a loop. On the (nexthopk,hopcounta) E route-list!}
other hand, loop cannot be formed if A accepts a duplicate
copy of the RREQ arriving via a trajectory that does not
The route update rule in Figure 3 is invoked whenever
a node receives a route advertisement. Lines (1) and (9)-
already include A.
(10) of the AOMDV route update rule ensure loop freedom.
In order to eliminate any possibility of loops, we main-
A proof is provided in the appendix. A key observation
tain a similar invariant as in the single path case. The major
here is that similar to AODV the following condition holds
difference, however, is that we accept and maintain multi-
good for two successive nodes i and j on any valid route to
ple next-hop routes as obtained by multiple route advertise-
ments, but we do this as long as the invariant is satisfied. destination d.
One trouble is that different routes to the same destination
( - s e q n u m t , advertised-hopcount!, i ) >
now may have different hopcounts. Therefore, a node must
be consistent regarding which one of these multiple hop- (-seqnum;, advertised-hopcount;, j),
counts is advertised to others. It cannot advertise different
hopcounts to different neighbors with the same destination where the comparison is in the lexicographic sense.
sequence number.
We build the AOMDV invariant based on a new notion of 3.2 Finding Link-disjoint Paths
“advertised hopcount.” The advertised hopcount of a node
i for a destination d represents the “maximum” hopcount of The above mechanism establishes multiple loop-free
the multiple paths for d available at i . “Maximum” hop- paths at every node. Here, we explore how we can make
count is considered, as then the advertised hopcount can these paths disjoint. Path disjointness has the nice property
never change for the same sequence number. The proto- that paths fail independently. There are two types of dis-
col only allows accepting alternate routes with lower hop- joint paths: node-disjoint and link-disjoint. Node-disjoint
counts. This invariance is necessary to guarantee loop free- paths d o not have any nodes in common, except the source
dom. and destination. In contrast, link-disjoint paths do not have

16
 if (seqnum: < seqnumjd) then (1)

seqnum! := seqnum!; (2)

if (i # d ) then (3)
advertised-hopcount! := CO; (4)
route-list! =NULL; (5)
insert ( j ,advertised-hopcount! + 1) into route-list!; (6)

else (7)
advertised-hopcount: := 0; (8)
endif
elseif (seqnum: = seqnum;) and (9)
((advertised-hopcount:, i ) > (advertised-hopcountjd, j ) ) then

insert ( j , advertised-hopcountjd + 1) into route-list:; (10)

endif

Figure 3. AOMDV route update rules. This is used whenever a node i receives a route advertisement to a destination
d from a neighbor j. The variables seqnum:, advertised-hopcount! and route-list: represent the sequence number,
advertisedhopcount and routelist for destination d at node i respectively.

any common link. Note that link-disjoint paths may have is I

common nodes.
We have investigated the issue of node- vs. link-
disjointness carefully. Generally speaking, using link-
disjointness is fine as long as links do fail independently.
This is often the case, except in situations where multiple
links f r o d t o a node fail together as the node moves out
of range. Of course, node-disjointness actually guarantees
that links fail independently. However, node-disjointness
is a much stricter condition than link-disjointness and thus
presents a much lesser number of disjoint routes in the sim- Figure 4. Illustration of link-disjoint path computa-
ulations we have run. This makes node-disjointness less tion in AOMDV. S , I and D denote the source, in-
effective. Thus, we stick to link-disjoint routes here. termediate node and the destination, respectively; is1
Also, here we are interested in investigating the fault and is2 are node-disjoint paths from I to S;id1 and
tolerance property of multipath routing, and not the load id2 are node-disjoint paths from I to D . There are
balancing property. In the latter case, use of node-disjoint two possible sets of link-disjoint paths between S and
routes will be more critical. With a simple modifica- D , is1 - i d l , is2 - id2 or is1 - id2, is2 - i d l .
tion, AOMDV can allow discovery of either node- or link-
disjoint routes and this can easily be controlled by a flag. So
the choice can be left to the implementor or even the user. and destination) is based.
Additional modifications are made in the route discovery
Property 1. Let a node SJlood a packet m in the network.
process to allow formation of node-disjoint paths from in-
The set of copies of m received at any node I (# S), each
termediate nodes to the source and destination. A collection
arriving via a different neighbor of S, defines a set of node-
of such node-disjoint paths forms a set of link-disjoint paths
disjoint paths from I to S.
between the source and destination. See Figure 4 for an il-
lustration. We first describe an important property of flood- Prooj We prove by contradiction. Suppose that the paths
ing on which the first part of the route discovery (forming taken by two copies of m received at I via different neigh-
node-disjoint paths from intermediate nodes to the source bors of S have a common node J . This implies that J must

17
 x copy of copy 1
via X

via Y
B

Figure 6. The second copy of RREQ via B is sup-

Figure 5. Illustration of Property 1 . S floods a packet pressed at intermediate node I. However two copies
in the network. A, B and C are neighbors of S. J trans- of the first copy (via A) still reaches the destination
mits only the first arriving copy of the packet (from A D. Both are replied to by D even though both carry
or B) and suppresses the latter. Two copies are re- the same first hop. The reverse paths will merge at
ceived at I, one via C and the other via A or B. Thus I I and then split again. But they will remain link dis-
finds two node-disjoint paths to S. joint.

have transmitted at least two copies of m, each received via ous than node-disjoint paths), the destination node adopts
a different neighbor of S . But in the flooding algorithm, a ‘‘looser’’ reply policy. It replies up to IC copies of RREQ
each node transmits the message at most once, a contradic- (arriving via unique neighbors), regardless of the firsthops
tion. 0
taken by these RREQs. Unique neighbors guarantee link
disjointness in the first hop of the RREP. Beyond the first
Observe that in the above property, the neighbors of S hop, the RREP follow the reverse routes that have been
uniquely identify the node-disjoint paths to S from any node set up already which are node-disjoint (and hence link-
I . Since the RREQs are flooded in AODV, replacing m disjoint). Each RREP arriving at an intermediate node takes
by RREQ and S by the source of the RREQ provides a a different reverse route when multiple routes are already
mechanism for an intermediate node ( I )to determine node- available. Note that because of the “looser” reply policy it
disjoint paths to the source. This concept is explained fur- is possible for the trajectories of RREPs to cross at an inter-
ther in Figure 5. mediate node (Figure 6).
Several changes are needed in the AODV route discov- The parameter k is used to control the number of RREPs
ery procedure to enable computation of link-disjoint paths
and thus to prevent a RREP storm. Also, our earlier ob-
between source-destination pairs. Each RREQ now carries
servation [21] indicated that additional routes beyond a few
an additional field called f irsthop to indicate the first hop
provide only marginal benefit, if any. We have used k = 3
(neighbor of the source) taken by it. Also, each node main-
in our experiments.
tains a firsthopdist for each RREQ to keep track of the
list of neighbors of the source through which a copy of the
RREQ has been received.
4 Performance Evaluation
At the intermediate nodes, unlike in AODV, duplicate
copies of RREQ are not immediately discarded. Each copy
is examined to see if it provides a new node-disjoint path to We have evaluated the performance of AOMDV with re-
the source . This is ascertained by examining the f irsthop spect to AODV using ns-2 simulations under a wide range
field in the RREQ copy and the f irsthop-list in the node of mobility and traffic scenarios. The goal is to address the
for the RREQ. If it does provide a new path, the AOMDV following questions:
route update rule (Figure 3) is invoked to check if a reverse
path can be set up. If a reverse path is set up and a valid
0 How does AOMDV compare with AODV, particularly
route to the destination is available at the intermediate node,
in terms of end-to-end delay and frequency of route
it sends back a RREP to the source. Just as in AODV, only discoveries, as node mobilities vary?
the first arriving RREQ copy is forwarded if a route to des-
tination is unavailable.
At the destination, reverse routes are set up just like in How does AOMDV compare with AODV with in-
the case of intermediate nodes. However, in the hope of crease in offered load (i.e., number of sessions a n d o r
getting link-disjoint paths (which would be more numer- packet rate per session) ?

18
4.1 Simulation Environment - this includes all possible delays caused by buffering dur-
ing route discovery, queuing delay at the interface, retrans-
We use a detailed simulation model based on ns-2 [SI. mission delays at the MAC, propagation and transfer times;
The Monarch research group in CMU developed support (iii) Route discoveryfrequency - the total number of route
for simulating multi-hop wireless networks complete with discoveries initiated per second; (iv) Normalized routing
physical, data link and MAC layer models [3] on ns- loud - the total number of routing packets “transmitted”
2. The distributed coordination function (DCF) of IEEE for each delivered data packet. Each hop-wise transmission
802.1 1 [ 121 for wireless LANs is used as the MAC layer. of these packets is counted as one transmission.
The radio model uses characteristics similar to a commer-
cial radio interface, Lucent’s WaveLAN [32]. WaveLAN is 4.2.2 Varying mobility
a shared-media radio with a nominal bit-rate of 2 Mb/sec
Figure 7 shows the four performance metrics as a function
and a nominal radio range of 250 meters. A detailed
of mobility. Max. speed of the nodes is varied from 0 m/s
description of simulation environment and the models is
to 30 m/s to change mobility. Max. speed of 0 d s corre-
available in [3, 81 and will not be presented here. Note
sponds to a static network. Average rate of link failures in
that the same simulation environment has been used before
our mobility scenarios increases between 0-50 per second
in several recent performance studies on ad hoc networks
as the max. speed increases between 0-30 m/s. The number
[3, 10, 13, 291. In our simulations, we use the latest AODV
of sessions and packet rate are fixed at 25 and 4 packets/sec,
specification [28]. Link layer feedback is used to detect link
respectively. Performance of AODV and AOMDV are sim-
failures.
ilar in the static case. Their performance differences, how-
Mobility and traffic models are similar to previously re-
ever, become more apparent at higher speeds (Figure 7). As
ported results using this simulator [3, 13, 291. The random
expected, the fraction of packets delivered] goes down for
waypoint model (31 is used to model mobility. Here, each
both, the ptorocols. However, AOMDV loses fewer packets
node starts its journey from a random location to a random
than AODV (3-5% less) in mobile cases.
destination with a randomly chosen speed (uniformly dis-
There is a tremendous reduction in the average end-to-
tributed between 0 and max. speed). Once the destination
end delay with AOMDV as shown in Figure 7 (b). Im-
is reached, another random destination is targeted after a
provement in delay is almost always more than 100%. This
pause. We consider only the continuous mobility case (i.e.,
is because availability of alternate routes on route failures
no pauses). To change mobility, we vary the max. speed of
eliminates route discovery latency that contributes to the
the nodes. A 100 node network in a field with dimensions
delay. Interestingly, for both protocols the delay increases
2200m x 600m is used.
with mobility only until the max. speed of IO m/s and
Traffic sources are CBR (continuous bit-rate). The
beyond that delays stabilize. With additional instrumenta-
source-destination pairs (sessions) are spread randomly tion, we found that packet drops at intermediate nodes due
over the network. Only 5 12 byte data packets are used. The to link failures beyond 10 m/s are dominated by packets
number of sessions and/or the packet sending rate in each with longer path lengths. In other words, the average hop-
pair are varied to change the offered load in the network. counts of delivered data packets comes down at high speeds.
All traffic sessions are established at random times near the
Thus, delays become insensitive to increase in mobility af-
beginning of the simulation run and they stay active until the ter a point as the packets delivered at high speeds are mostly
end. Simulations are run for 500 simulated seconds. Each those that travel along shorter paths. Frequency of route dis-
data point represents an average of ten runs with identical
coveries and the routing load behave similarly with varying
traffic models, but different randomly generated mobility mobility ( Figure 7(c) and (d)). As expected, AOMDV per-
scenarios. Each plot shows error bars indicating the 90% forms better in both the metrics with improvements staying
confidence interval. Identical mobility and traffic scenarios around 20%.
are used across all protocol variations.
4.2.3 Varying offered load
4.2 Simulation Results
We first vary the number of sessions from 10-50 with packet
4.2.1 Performance Metrics rate per session fixed at 2 packetslsec (Figure 8). We keep
the max. speed constant at 10 m/s in this set of experiments.
We evaluate four key performance metrics: (i) Packet de- Figure 8 (a) plots the average delay (in seconds) against the
liveqfruction - ratio of the data packets delivered to the throughput (in Kb/s). Note the significant reduction in de-
destination to those generated by the CBR sources; or a re- lay for AOMDV for the same throughput - hallmark of a
lated metric received throughput in Kb/sec received at the good routing protocol [2]. Reduction in routing load is also
destination; (ii) Average end-to-end delay of data packets observed (Figure 8 (b)) as before.

19
 I , I
AODV
AOMDV
-
0.6
0.55 - AOMDV
AODV -
1 - - --.
-3
C 0.95

0.9 I 0.4 -
.e- 2 0.35 - ,
5
- 0.85 0.3 - X

-4
B ..... .I .
D
e 0.8 I ,
:

0.15 - 4.............,,A.___......-..;.,.......----?
0.75 0.1 r ;
0 5 10 15 20 25 30
0.05 I I
Max. Speed (ds) 50 100 I50 200 250 300 350
Throughpul (Kbls)
(a) Packet Delivery
(a) Delay vs. Throughput

0.6
15
0.5 AODV -
-p 0.4
14 AOMDV

a6 0.3
2 0.2

0.I

n
0 5 10 15 20 25 30
Max. Speed (IPA)

IO 15 20 25 30 35 40 45 50
(b) End-to-End Delay
#Sessions

(b) Routing Load

Figure 8. Performance with varying number of ses-

sions.

Figure 9 studies the performance with varying packet

o 5 IO 15 20 25 30
Max. Speed ( m / s )
rate. We vary the packet rate at each source from 0.5-8
packetdsec while keeping the max. speed and the number
(c) Route Discovery Frequency of sessions fixed at 10 m/s and 25, respectively. As before,
AOMDV provides much lower delay for the same through-
11
put (Figure 9(a)). Note the initial sharp reduction in normal-
in - AODV - ized routing load with increase in packet rate (Figure 9(b)).
At very low packet rates, most routes become invalid by
the time they are used again. This is regardless how many
routes are computed each time. This is because link failure
rates are very high compared to packet rate. Thus, multi-
path and single path protocols perform alike at low packet
rates. Both protocols start performing significantly well -
with AOMDV relatively better - as the packet rate is in-
0 5 IO IS 20 25 30
Max. Speed (ds) creased beyond the rate of link failures. At higher packet
rates, however, the routing overhead in both protocols is
(d) Routing Load
dominated by route retry attempts as it takes longer to get
a route due to the high network load. Relatively, AOMDV
is affected more since it also has the additional overhead of
Figure 7. Performance with varying mobility. more RREPs per route discovery. But notice that at these

20
 shortest pairs of node- (link-) disjoint paths. Later, an im-
0.9 proved technique was proposed by Sidhu et al. [311 to com-
pute node-disjoint paths. Another well known example of
a multipath routing algorithm is the OSPF [ 191, a link state
0.6
protocol that computes multiple paths of equal cost. More
recently, multipath distance vector algorithms that use dif-
0.4
fusing computations to construct and maintain DAGS have
0.2
been proposed [33, 351. However, all the above algorithms
0. I have high overheads that make them inefficient for band-
01 ' " " " " width limited wireless networks. They are designed to work
0 50 100 150 200 250 300 350 400 450
in the framework of proactive protocols, prevalent in the In-
Throughput (Kbk)
ternet, where overheads are not such major concern.
(a) Delay vs. Throughput There has been some interest in the ad hoc networking
community to employ multipath routing algorithms. Two of
24
the on-demand protocols, DSR [ 141 and TORA [24] have
AODV
AOMDV
-
........ . built-in capability to compute multiple paths. But either
of them suffers from a different set of performance prob-
lems. DSR uses source routing, by virtue of which it can
detect loops easily and can gather a lot of routing informa-
tion per route discovery. However, aggressive use of route
caching, lack of effective mechanisms to purge stale routes
and cache pollution leads to problems such as stale caches
and reply storms. These problems not just limit the per-
0 1 2 3 4 5 6 7 8 formance benefits of caching multiple paths, they can even
Packet rate (packetskec) hurt performance in many cases [IO, 291. These problems
(b) Routing Load are, however, being addressed [ l 1, 171. TORA [14], on the
other hand, builds and maintains multiple loop free paths
without use of source routing. It also can detect network
partitions. TORA uses an idea based on link reversals [9] to
Figure 9. Performance with varying packet rates. recover from link failures. Performance studies have shown
that TORA suffers from high overheads primarily because
high rates, the network is already under saturation (almost of the requirement of reliable, in-order delivery of routing
40% packets are dropped by both protocols), and in an over- control messages [3].
all sense, AOMDV still performs better as it gives lower Routing On-demand Acyclic Multipath (ROAM) [30] is
end-to-end delay. an on-demand, multipath distance vector algorithm based
on diffusing computations. Like TORA, ROAM can also
5 Related Work detect network partitions. But on the downside, state infor-
mation must be maintained at each node during route dis-
Multipath routing and its applications [4, 181 have been covery; this requires close coordination between nodes in-
well studied in the networking literature, in particular for creasing overheads. Thus ROAM is better suited for static
wired networks. In a broad sense, multipath routing en- ad hoc networks or networks with low node mobility.
ables fault tolerance and also facilitates load balancing. An A technique is proposed in [ 151 to allow AODV to main-
early work by Maxemchuk [ 181 on an application of mul- tain backup routes at the neighboring nodes of a primary
tipath routing known as dispersity routing discusses how route. This can be done with no additional overheads. The
a message can be dispersed along multiple paths by split- idea is to avoid dropping packets in flight when the primary
ting it in order to achieve smaller average delay and delay route fails. However, every node still has at most one route
variance. Since then, there has been a significant amount per destination just as in AODV. Finally, none of the above
of work done on multipath routing for both connection- mentioned protocols explicitly take path disjointness into
oriented (e.g., [ 1,6]) and connection-less technologies (e.g., account.
[20, 351). Path disjointness has been considered in [2 1,25, 16,341,
Distributed protocols to compute multiple disjoint and but all of them use source routing. Nasipuri et al. [21] pro-
loop-free paths are more related to our interests. Ogier pose extensions to DSR to compute multiple disjoint paths
and Shacham [23] describe a distributed algorithm to find for overhead reduction in mobile networks. They study the

21
 effect of number of multiple paths, path lengths on rout- 0096264. Mahesh Marina is supported by an OBR comput-
ing performance using analytical modeling and packet-level ing research award in the ECECS department, University of
simulations. Split Multipath Routing (SMR) [ 161 is another Cincinnati.
disjoint multipath protocol using source routing. SMR is
similar to multipath DSR [21] except that the former uses Appendix
a modified flooding algorithm and the data traffic is split
among the multiple paths. Pearlman et al. [25] analyze
Theorem 1. The route update rule in AOMDV (Figure 3 )
the performance impacts of alternative path routing for loud
yields loop free routes.
balancing. They use a technique called diversity injection
to compute node-disjoint paths. In [34], an improvement P r o o j Suppose that a loop of size m, (il,i2, ..,im,il)
over the diversity injection technique [25] is proposed to forms in a route to a destination d. Note that nodes i and
find more node-disjoint paths. j in the code are two consecutive nodes in the route, and

6 Conclusions and Future Work seqnumf 5 seqnum;.

Therefore, the following must be true among the nodes in
Multipath routing can be used in on-demand protocols
the loop so formed.
to achieve faster and efficient recovery from route failures
in highly dynamic ad hoc networks. In this paper, we have seqnumfl 5 seqnumf2 5 .. 5 seqnumfm 5 s e q n u m f l ,
proposed an on-demand, multipath distance vector proto-
col AOMDV that extends the single path AODV protocol to which implies
compute multiple paths. There are two main contributions
of this work: seqnumfl = seqnumf2 = . . . = seqnumtm = seqnumfl
1. We use the notion of an advertised hopcount to main- This in turn implies the following condition holds in line 9.
tain multiple loop-free paths at each node.
(advertised-hopcountfl,i l ) >
2. We show how route discovery mechanism in the
AODV protocol can be modified to obtain link-disjoint > .. >
(advertised-hopcountf2,i 2 )
multiple paths from source and intermediate nodes to (advertised-hopcounttm i m ) >
the destination.
(advertised-hopcountfl i l ) .
We have studied the performance of AOMDV relative to
AODV under a wide range of mobility and traffic scenarios. Then,
We observe that AOMDV offers a significant reduction in
delay, often more than a factor of two. It also provides up to (advertised_hopco2Lnt,4,i l ) >
about 20% reduction in the routing load and the frequency (advertised-hopcountfl,i l ) ,
of route discoveries. In general, AOMDV always offers a
superior overall routing performance than AODV in a vari- which clearly is impossible. Thus, routes formed by
ety of mobility and traffic conditions. AOMDV are loop free. 0
We are currently working on augmenting the AOMDV
protocol with a technique to uniquely identify each disjoint References
path on an end-to-end basis. This feature is very useful
when it is necessary to route always along one specific path.
[ I ] S. Bahk and M. E. Zarki. Dynamic Multi-path Routing and
For example, TCP retransmission timeout computation is How it Compares with other Dynamic Routing Algorithms
sensitive to message reordering which can be avoided by for High Speed Wide Area Networks. In Proceedings ofrhe
always following the same end-to-end path as long as it is ACM SIGCOMM, pages 53-64, 1992.
available. We will also look into other issues related to on- 123 D. Bertsekas and R. Gallager. Data Networks. Prentice-Hall,
demand multipath routing - for example, the availability 1992.
of multiple paths in relationship with node density and load [3] J. Broch, D. Maltz, D. Johnson, Y.-C. Hu, and J. Jetcheva.
balancing with multiple paths. A Performance Comparison of Multi-Hop Wireless Ad
Hoc Network Routing Protocols. In Proceedings of the
IEEElACM MOBICOM, pages 85-97, 1998.
Acknowledgments [4] J. W. Byers, M. Luby, M. Mitzenmacher, and A. Rege. A
Digital Fountain Approach to Reliable Distribution of Bulk
This work is partially supported by NSF CAREER grant Data. In Proceedings of the ACM SIGCOMM, pages 5 6 6 7 ,
ACI-0096 186 and NSF networking research grant ANI- 1998.

22
 [5] R. Castaneda and S. R. Das. Query Localization Techniques [231 R. Ogier, V. Rutenburg, and N. Shacham. Distributed Algo-
for On-demand Protocols in Ad Hoc Networks. In Proceed- rithms for Computing Shortest Pairs of Disjoint Paths. IEEE
ings of the IEEEIACM MOBICOM, pages 186-194, 1999. Transactions on Information Theory, 39(2):443-4.55, 1993.
[6] I. Cidon, R. Rom, and Y. Shavitt. Analysis of Multi-Path [24] V. D. Park and M. S. Corson. A Highly Adaptive Distributed
Routing. IEEE Transactions on Networking, 7(6):885-896, Routing Algorithm for Mobile Wireless Networks. In Pro-
1999. ceedings of the IEEE INFOCOM, pages 1405-1413, 1997.
[7] S. R. Das, R. Castaneda, and J. Yan. Simulation-based Per- [25] M. R. Pearlman, Z. J. Haas, P. Sholander, and S. S. Tabrizi.
formance Evaluation of Routing Protocols for Mobile Ad On the Impact of Alternate Path Routing for Load Balancing
hoc Networks. ACMIBaltzer Mobile Networks and Applica- in Mobile Ad Hoc Networks. In Proceedings of the ACM
tions (MONET),5(3): 179-189, 2000. MobiHoc, pages 3 -10,2000.
(81 K. Fall and K. Varadhan(Eds.). ns notes and documentation, [26] C. E. Perkins and P. Bhagwat. Highly Dynamic Destination-
1999. available from http://www-mash.cs.berkeley.edu/ns/. Sequenced Distance-Vector Routing (DSDV) for Mobile
[9] E. Gafni and D. Bertsekas. Distributed Algorithms for Computers. In Proceedings of the ACM SIGCOMM, pages
Generating Loop-free Routes in Networks with Frequently 234-244, 1994.
Changing Topology. IEEE Transactions on Communica- [27] C. E. Perkins and E. M. Royer. Ad Hoc On-Demand Dis-
tions, 29(1):11-18, 1981. tance Vector Routing. In Proceedings of the IEEE Workshop
[IO] G. Holland and N. H. Vaidya. Analysis of TCP Perfor- on Mobile Computing Systems and Applications (WMCSA),
mance over Mobile Ad Hoc Networks. In Proceedings of pages 90-100, 1999.
the IEEE/ACM MOBICOM, pages 219-230, 1999. [28] C. E. Perkins, E. M. Royer, and S. R. Das. Ad
[ 1 I] Y.-C. Hu and D. Johnson. Caching strategies in On-demand Hoc On Demand Distance Vector (AODV) Routing.
Routing Protocols for Wireless Ad Hoc Networks. In Pro- http://www.ietf.org/internet-drafts/draft-ietf-manet-aodv-
ceedings of the IEEE/ACM MOBICOM, pages 23 1-242, 07.txt, Nov 2000. IETF Internet Draft (work in progress).
2000. [29] C. E. Perkins, E. M. Royer, S . R. Das, and M. K. Marina.
[I21 IEEE Standards Department. Wireless LAN medium ac- Performance Comparison of Two On-demand Routing Pro-
cess control (MAC) and physical layer (PHY) specifications, tocols for Ad Hoc Networks. IEEE Personal Communica-
IEEE standard 802.1 1-1997. tions, 8(1):16-28, 2001.
[13] P. Johansson, T. Larsson, N. Hedman, and B. Mielczarek. [30] J. Raju and J. J. Garcia-Luna-Aceves. A New Approach to
Routing Protocols for Mobile Ad-hoc Networks - A Com- On-demand Loop-Free Multipath Routing. In Proceedings
parative Performance Analysis. In Proceedings of the of the Int’l Con$ on Computer Communications and Net-
IEEE/ACM MOBICOM, pages 195-206, 1999. works (IC3N),pages 522-521, 1999.
[ 141 D. Johnson and D. Maltz. Dynamic Source Routing in Ad
Hoc Wireless Networks. In T.Imielinski and H. Korth, edi- [31] D. Sidhu, R. Nair, and S. Abdallah. Finding Disjoint Paths
in Networks. In Proceedings of the ACM SIGCOMM, pages
tors, Mobile computing, chapter 5. Kluwer Academic, 1996. 43-51, 1991.
[I51 S. J. Lee and M. Gerla. AODV-BR: Backup Routing in Ad
hoc Networks. In Proceedings of the IEEE Wireless Commu- [32] B. Tuch. Development of WaveLAN, an ISM band wireless
LAN. AT&T Technical Journal, 72(4):27-33, 1993.
nications and Networking Conference(WCNC),pages 1311-
[33] S. Vutukury and J. J. Garcia-Luna-Aceves. MDVA: A
1316,2000.
Distance-Vector Multipath Routing Protocol. In Proceed-
[I61 S. J. Lee and M. Gerla. Split Multipath Routing with Max-
ings of the IEEE INFOCOM, pages 5.57-564, 2001.
imally Disjoint Paths in Ad Hoc Networks. In Proceedings
of the IEEE ICC, pages 3201-3205,2001.
[34] K. Wu and J. Harms. Performance Study of a Multipath
[I71 M. K. Manna and S. R. Das. Performance of Route Routing Method for Wireless Mobile Ad Hoc Networks.
Caching Strategies in Dynamic Source Routing. In Proceed- In Proceedings of the IEEE Int ’1 Symposium on Modeling,
ings of the Int’l Workshop on Wireless Networks and Mo- Analysis and Simulation of Computer and Telecommunica-
bile Computing ( WNMC) in conjunction with Int ’1 Cont on tion Systems (MASCOTS),pages 99-107,2001.
Distributed Computing Systems (ICDCS), pages 42.5-432, [35] W. Zaumen and J. J. Garcia-Luna-Aceves. Shortest multi-
2001. path routing using generalized diffusing computations. In
[I81 N. F. Maxemchuk. Dispersity Routing. In Proceedings of Proceedings of the IEEE INFOCOM, pages 1408-1417,
the IEEE ICC, pages 41:1O-41:13, 1975. 1998.
[I91 J. Moy. OSPF version 2. RFC 1247, 1991.
[20] S. Murthy and J. J. Garcia-Luna-Aceves. Congestion-
Oriented Shortest Multipath Routing. In Proceedings ofthe
IEEE INFOCOM, pages 1028-1036, 1996.
[21] A. Nasipuri, R. Castaneda, and S. R. Das. Performance of
Multipath Routing for On-demand Protocols in Mobile Ad
Hoc Networks. ACM/Kluwer Mobile Networks and Appli-
cations (MONET),6(4):339-349, 2001.
[22] S.-Y. Ni, Y.-C. Tseng, Y.-S. Chen, and J.-P. Sheu. The
Broadcast Storm Problem in a Mobile Ad Hoc Network. In
Proceedings of the IEEE/ACM MOBICOM, pages 151-1 62,
1999.

23