Message Ferrying: Proactive Routing in Highly-partitioned Wireless Ad Hoc

Networks

Wenrui Zhao and Mostafa H. Ammar

Networking and Telecommunication Group
College of Computing, Georgia Institute of Technology

Atlanta, Georgia 30332
wrzhao, ammar @cc.gatech.edu

Abstract physical factors, might preclude nodes from communicat-

ing with others and keep the network in partitioned state.
An ad hoc network allows devices with wireless inter- For example, in a battlefield, nodes equipped with short
faces to communicate with each other without any pre- range radio may move out of range of others. In addition,
installed infrastructure. Due to node mobility, limited ra- the wide physical range of the deployed area in some cir-
dio power, node failure and wide deployment area, ad hoc cumstances will also prevent full coverage due to the cost
networks are often vulnerable to network partitioning. A involved. Under these environments, most existing rout-
number of examples are in battlefield, disaster recovery and ing algorithms will fail to deliver messages to their destina-
wide area surveillance. Unfortunately, most existing ad hoc tions since no route is found due to network partition. This
routing protocols will fail to deliver messages under these raises the question of how to deliver data in a constantly
circumstances since no route to the destination exists. In disconnected network. While certain applications, like real-
this work, we propose the Message Ferrying or MF scheme time video or audio streaming, require bandwidth or delay
that provides efficient data delivery in disconnected ad hoc guarantees to be useful, other applications would benefit
networks. In the MF scheme, nodes move proactively to from the eventual and timely delivery of data, even in the
send or receive messages. By introducing non-randomness presence of network partition. A number of these applica-
in a node’s proactive movement and exploiting such non- tions include messaging, file transfer, email and other non
randomness to deliver messages, the MF scheme improves real-time applications. We expect that this kind of highly-
data delivery performance in a disconnected network. In partitioned ad hoc networks would not be uncommon in
this paper, we propose the basic design of the MF scheme practice, e.g. in battlefield, disaster recovery and wide area
and develop a general framework to classify variations of surveillance. Thus, efficient routing algorithms must be de-
MF systems. We also study ferry route design problem in veloped to route messages to their destinations despite of
stationary node case which is shown to be NP-hard and pro- network partition.
vide an efficient algorithm to compute ferry route. The existence of network partitioning requires a new ap-
proach other than the traditional “store-and-forward” rout-
ing paradigm used in most current ad hoc routing algo-
1 Introduction rithms, in which messages are dropped if no route is found
to reach the destinations. Instead, a device should buffer and
carry the messages until it has a chance to forward them.
An ad hoc network allows devices with wireless inter-
Thus, devices need to store, carry and forward messages in
faces to communicate with each other without any pre-
highly-partitioned ad hoc networks [15, 2].
installed infrastructure. Many routing algorithms have been
developed for ad hoc networks [5, 7, 13, 11, 12, 14, 4]. Several schemes have been proposed to solve the rout-
However, most of these algorithms are designed for fully ing problem in highly-partitioned wireless ad hoc networks
connected ad hoc networks assuming persistent end-to-end [15, 2, 8]. Epidemic routing [15] is a flooding-style algo-
connectivity between any two nodes. While it is desirable rithm in which nodes forward messages to other nodes they
to maintain a connected network, ad hoc networks are of- meet. This scheme will deliver all messages given unlimited
ten vulnerable to network partitioning. Node mobility, lim- time and memory. The blind-flooding of epidemic routing
ited radio range, physical obstacles, severe weather or other transmits a lot of redundant messages and requires a large
amount of buffering, which leads to poor scalability. The tacking each of these subproblems we develop an efficient
work in [2] proposed an improved scheme over epidemic algorithm to compute ferry routes.
routing by exploiting node mobility statistics. Nodes esti- The rest of this paper is organized as follows. We intro-
mate the probabilities of meeting other nodes in the future duce the Message Ferrying scheme in Section 2. In Section
and forward messages based on these probabilities. It can 3 we describe the ferry route design problem in the station-
deliver more messages, as compared to epidemic routing, in ary node case and present an algorithm to compute ferry
networks with limited memory. Both schemes in [15] and routes. Section 4 presents our preliminary simulation re-
[2] are reactive in that when disconnected, nodes passively sults. We conclude the paper in Section 5.
wait for their chances to re-connect, e.g. meet with other
nodes, which may lead to significant and even unacceptable
transmission delays and low throughput. To address this 2 Message Ferrying Scheme
problem, the work in [8] proposed a proactive scheme in
which mobile nodes actively modify their trajectories in or- The Message Ferrying (MF) scheme is a proactive ap-
der to transmit messages as soon as possible. The proposed proach for routing in disconnected ad hoc networks. It
scheme computes a trajectory for sending messages through addresses the disconnection problem by introducing non-
intermediate nodes to reach the destinations and optimizes randomness to node mobility and exploiting such non-
both the transmission delay and the trajectory modification randomness to provide connectivity. In the MF scheme, the
for the case where a single message is transmitted in the net- network devices are classified as message ferries (ferries for
work. However, the work in [8] assumes that nodes have short) or regular nodes based on their roles in communica-
full knowledge of the location and movement of all nodes tion. Ferries are devices which take responsibility of car-
in the network, and cannot support parallel transmission of rying messages between disconnected nodes, while regular
multiple messages in the network. These assumptions make nodes are devices without such responsibility. Ferries move
it difficult for use in realistic networks. around the deployed area according to known routes, col-
In this work, we propose a new proactive routing scheme lect messages from the sending nodes and deliver messages
for disconnected wireless ad hoc networks. The new to their destinations or other ferries. With knowledge about
scheme is referred to as Message Ferrying, inspired from ferry routes, nodes can adapt their trajectories to meet the
its real life analog. In this scheme, nodes move proactively ferries and transmit or receive messages. Ferries actually
in order to send or receive messages. The main idea of the serve as “rendezvous points” for message senders and re-
MF scheme is to introduce non-randomness in the proac- ceivers. By using ferries as relays, nodes can communicate
tive movement of nodes and exploit such non-randomness with others that are disconnected.
to help deliver messages. In the MF scheme, a set of de- Figure 1 shows an example of how message transfer
vices called message ferries (or ferries in short) take re- is done using a single ferry in a mobile ad hoc network.
sponsibility for carrying messages between disconnected The ferry moves on a known route which is depicted using
nodes. Message ferries move around the deployed area ac- dashed line in Figure 1. In Figure 1(A), the sending node S
cording to known routes and communicate with other nodes actively approaches the ferry and forwards its messages to
they meet. With knowledge of the ferry routes, nodes can the ferry which will be responsible for delivery. In Figure
adapt their trajectories to meet the ferries and transmit or re- 1(B), the receiving node R communicates with the ferry and
ceive messages. By using ferries as relays, nodes can com- receives its messages.
municate asynchronously with other nodes that are discon-
nected. The MF scheme differs significantly from previous 2.1 Functions and Capabilities of MF Scheme
schemes in its non-random proactive movement of nodes. Components
With proactive movement of both the ferries and the nodes,
the MF scheme provides regular connectivity in an other- As discussed above, message ferrying is a broad concept,
wise disconnected ad hoc network. with the potential for many variations in specific design and
In this paper, we propose the basic design of the MF implementation. We develop a general framework for MF
scheme and develop a general framework to classify vari- systems based on the capabilities and functions performed
ations of MF systems. We then study the ferry route de- by the various components of a message ferrying system.
sign problem for the case where nodes are stationary. Since We consider the following five dimensions that define the
the MF scheme depends on node movement to deliver mes- context for a Message Ferrying scheme:
sages, the choice of ferry routes will have significant impact
on the achieved throughput and delay. We formulate this 1. Ferry Mobility: The movement of the ferry is a critical
problem as an optimization problem and divide it into two feature of an MF scheme. We consider two possible
subproblems, one of which is shown to be NP-hard. By at- cases, which may co-exist in the same scenario:
 case, a message might be forwarded through multiple
S ferries while being routed to its destination.

R 4. Level of Regular Node Coordination: The regular

nodes can use the ferry to deliver data in two funda-
mentally different ways:

Ferry
 The regular nodes operate independently, with
each regular node acting on its own and in charge

(A) S approaches Ferry and transmits messages

 of delivering to and receiving from the ferry.
The regular nodes coordinate with each other to
form connected clusters. Within a cluster, one or
more gateway nodes are in charge of communi-
R cating with the ferry. The other nodes commu-
Ferry nicate with these gateway nodes using traditional
ad-hoc network routing protocols.

5. Ferry Designation: Ferries can be either specially des-

ignated nodes or regular nodes that are (perhaps tem-
S porarily) elevated to perform ferry functions. In the
former case, it may be possible to assume that a ferry’s
resources (power, memory, disk storage) are not as
limited as typical nodes. For the latter case, there is, of
(B) R approaches Ferry and receives messages
course, the question of when and how to change node
designation.
Figure 1. An example of message delivery us-
ing a ferry as relay. In this paper, we focus on the case where regular nodes
are stationary and a single ferry is used. The study of other

 The ferry mobility is determined for non-

variations of message ferrying, such as in mobile networks
and the use of multiple ferries, is deferred to future research.
messaging reasons: For example, when piggy-
backing a ferry on a metropolitan area bus, the 3 Ferry Route Design for Stationary Node
route the ferry takes is determined based on
Case
passenger-carrying concerns and not message-

 forwarding concerns.
The ferry mobility is specifically designed for
We describe the basic design of the MF scheme in the
previous section. In this section, we focus on the ferry route
improving the performance of messaging. For design problem in the stationary node case. In this case,
example, the ferry is implemented in a subset of nodes are located at fixed positions and network partition
robots dispersed in a disaster area, and the mobil- prevents some nodes from communicating with others. The
ity of the ferry robots is specifically optimized for ferries are used to relay messages between disconnected
maximizing the efficiency of messaging among nodes. A basic question is how to design good or optimal
the other robots. ferry routes, given node positions and their communication
requirements. In this work, we assume that a single ferry is
2. Regular Node Mobility: While the ferry is always a
deployed which moves at a constant speed and the expected
mobile entity, the regular nodes can be stationary or
mobile. Similar to ferry mobility, mobile regular nodes
can move for non-messaging reasons or specifically to
Let   
traffic between nodes is known in advance.

be a set of nodes that want

improve messaging performance.
3. Number of Ferries and Level of Coordination: In gen-

to communicate. The ferry moves at a constant speed .
Let be the transmission range of the nodes and the ferry1 .
The nodes and the ferry communicate via a wireless chan-
eral, an MF system may have multiple ferries, each nel when they are within range of each other. Assuming
with a possibly different set of capabilities. Ferries 1 The results and algorithm developed in this section can be easily ex-
may operate completely independently of each other tended to handle the case where the nodes and the ferry have different
or their movements may be coordinated. In this latter transmission ranges.
that communication between the nodes and the ferry uses C
such that visits all nodes in  and the average delay aF
different radio frequencies in each direction, sending and
receiving can take place simultaneously. Let be the trans-
mission rate between the nodes and the ferry. When the
 is minimized.

Property 1. If >C +@?b\>?+ c d

for all and , then OFe T FU T
ferry is in range of multiple nodes, some policy is used to for any route .
schedule the transmission and reception of nodes. Instead
2. If route gC f
is the reverse route of , that is in the C C Cgf
of focusing on specific MAC scheduling policies, we de-

 #"%$ &'() 
fine a scheduling policy in a more general sense. Let be  G F Y GFihj TU F T
ferry visits the nodes in the reverse order as in , then
OF NFXh

 
  !
 * T FU T
the power set of . Then a policy can be defined as a . Thus either or is no more
function
+ -,.+*0/ +. For a set of nodes and a than

C
.

 k
node , defines the portion of transmission time

6 +8 * 79 -,:+;*0/<=( -,.+*0/12& +453 *

allocated to node when the ferry is within range of nodes 3. Let be a shortest route that visits all nodes in , e.g.
in . Obviously, we have if and an optimal traveling salesman tour for . If is an

> +@? .
6 +   ? nodes , then  TlUF nT m No m T FU T
optimal route for any MFR-delay problem defined on
.
Let be the expected traffic from node to , mea-
A += ?79 >B+@?
 +  6 ? 79  + > ?+
sured in bits per second. Define and Proof:
, which are the total incoming and outgoing
C GF+H? Y N?F +  T FUpT Y [ U1]^` _ Y [ _:` ]U  TU F T
 + D CED  ? GF+H? C GF OF4 T FU T
traffic for respectively. For any ferry route , we denote 1. . So, by the definition
its length as . Let be the average delay for traffic of , it is easy to prove that .

GF+H? Y N+@FX? h  TlUF gT Y [ Ub]@` _ Y T lUFXqh T Y [ ]@`U _ h  TU F T

from node to in route . Then the average delay for
all traffic can be defined as
6
 F  6 JIKJ+.IKL ?+.IML ?IM> +@>? +@N? F+@? D CEDrsD CtfuD V:+@F ? Y V:+HFX? h vD CED GF Y NFXh TU F T
2. We have since

GF T F GT FXh OF GFih

and . Thus .

U
Both and are positive, so either or is no
. more than .

C > +@? GF Cgf GFXh D CECD wr k

The ferry route problem consists of finding an optimal

OF x
route such that the bandwidth requirements are met 3. Let be the reverse route of . By Property 2, either

 o m T FU T  o+H?  TlUo T Y [ U]@_ y TJUF T Y [ U]@` _ y TJUF T

and the average delay is minimized. Rather than ad- or is not more than . Since is optimal,

 o y TlUF T TlUF T m  o m T FU T
dressing the combined problem, we break it into two sub- . . Thus
problems. The first one seeks to find a route that mini- . Therefore, we have .
mizes the average delay for the expected traffic matrix with-
out considering the bandwidth requirements. The second
sub-problem extends the route generated in the first sub-
problem, if necessary, to meet the bandwidth requirements. Theorem 1 The MFR-delay problem is NP-hard.

3.1 Delay problem Proof: Property 1 shows that when the traffic matrix
is symmetric, the optimal route for the MFR-delay problem
has shortest length. So we can reduce the Euclidean Trav-

to node K?
delay. The average delay for messages from node
consists of the waiting time at
OF+@?
In the first sub-problem, we only consider the message

before PQ+H? R+H? + +

eling Salesman Problem to the MFR-delay problem. Since
the Euclidean Traveling Salesman Problem is NP-complete
[10], we prove that the MFR-delay problem is NP-hard.
K ? F
transmitted to the ferry, and the carrying time at the ferry
The above Property and Theorem 1 relate the MFR-delay

PS+@? R+@? TJU T

before delivered to . With the assumption of constant bit
problem to the traveling salesman roblem (TSP) and lead
rate traffic, equals to . For simplicity, we measure us to adapt solutions from the well-studied TSP. We adopt a

 + K?
the carrying time
WV H+F ?
as the time for the ferry to move from two-phase approach to solve the MFR-delay problem. First

 + X? C O+HF ?  TJUF ZT Y\[^] _W`U

’s position to ’s position. Let be the distance from a starting route is generated using some TSP algorithm, e.g.
nodes to in route . Then . The nearest neighbor, greedy, or LKH [1, 6]. Since TSP al-
delay sub-problem, denoted as MFR-delay problem, can be gorithms only try to optimize the route length, instead of
defined as follows. the message delay, we can further reduce the average de-

and >+H?
Definition (MFR-delay problem) Let be a set of nodes
be the expected traffic from node to node , the
 + K? C
lay of the route by applying delay-based local optimization
techniques similar to the 2-opt and 2H-opt techniques [1] in
TSP. These delay-based techniques, denoted as 2-opt-delay
MFR-delay problem consists of finding an optimal route and 2H-opt-delay respectively, work as follows.
  j{ | *gz * {
z| , 6#?Œ 79 M+@?6 ‹a?Œ Y „a+ /; y A +
2-opt-delay. A 2-opt swap removes two edges and
from the route and replaces them with edges
and while maintaining the route. This technique
tries to reduce the delay of the route using 2-opt swaps T FMUŽ T Y ?79 ‹ ?
until no better route can be found.
 2H-opt-delay. A 2H-opt swap moves a node from one
Similarly we can derive constraints on the receiving di-
rection. After transformation, we get the following linear
programming problem which can be solved efficiently us-
position in the sequence of route to another. This tech-

Œ ‹ ?
nique tries to reduce the delay of the route using 2-opt ing methods like Simplex [9].
and 2H-opt swaps until no better route can be found.
minimize
?Œ79 (1)

 ,u +H?S
The relation between route length and route delay shown
in the above Property justifies our use of TSP algorithms A + /‘‹’? y A“+lXD C0 D  „ +
in our approach. The following theorem provides an upper subject to
?Œ79
bound for the delay obtained by this approach.
 ,u +HŠ ?S   + /‘‹a? y  + XD Cj rD  „ +Š
C k }
Theorem 2 Let is an optimal route for the MFR-delay
problem. Let be the route generated by a TSP algorithm ?79
& m ‹? ( m c m M
C  F m~ }9 o
with approximation ratio . Then the resulting delay of
route satisfies .
and
By combining the solutions for the two sub-problems,
Ctf a F  o m y D CET DFXm h T
we now present the algorithm, called MFR algorithm, for
Proof: Let be an optimal TSP tour. Then computing ferry routes which is shown in Figure 2. The
}€D C f D mn~ }9Go
by Property 3. Thus, from Property 2, we have
.
performance of MFR Algorithm in disjointed node case can
be established as follows.

3.2 Bandwidth problem MFR Algorithm

Cj 1. Compute a TSP tour Cj

for nodes in  using a TSP
Given the route
we now consider how to extend C‚
generated in the first sub-problem,
to meet a node’s band-
approximation algorithm.

of a node is } }
width requirement. For any route, the achieved bandwidth
where is the fraction of the ferry route
in which the node is in transmission and is the transmis-  C0
2. Apply 2H-opt-delay local optimization technique to
to reduce the average delay.

• ‹M ‹iN ‹ ŒE– Cj ” 

sion rate between the ferry and the node. Thus to increase a 3. Generate the linear programming problem as in For-
node’s bandwidth we need to increase its transmission time mula 1 for route and solve it. Let
be the optimal solution.
Cj
which is also affected by the scheduling policy used. Since

‹ + c
increasing a node’s transmission time leads to longer ferry 4. Modify to meet each node’s bandwidth require-
route and larger delay for other traffic, we should minimize ment by taking a detour of length on segment .
the amount of route extension.
We formulate the bandwidth sub-problem as a linear

pose jC 
programming (LP) problem as follows. We first decom-
into segments by cutting Cj
when it enters or
Figure 2. Algorithm for computing ferry route.

( ~ B+ ƒ CjO„ + 
leaves a node’s coverage area and index these segments as

!†ƒ ‡ˆ } C
. Let be the allocated transmission time for

M+@? ‡‰+
node in route . Define a matrix such that Theorem 3 Let be the route generated by the MFR Al-

 k
the entry of is the portion of transmission time al- gorithm and be the approximation ratio of the TSP algo-

d
located to node when the ferry is moving on segment rithm used in the MFR Algorithm. Route is an optimal

the variable
‹ + c „O(+Š m c m ƒ
and the matrix ‡‰Š
according to the scheduling policy. Similarly, we define
for the receiving direc- 
route for the ferry route problem and is the speed of the
ferry. Let be the transmission range of the nodes and the
~
+
tion2 . Let , , be the extra time the ferry spends ferry. If the distance between any two nodes is at least ,
e.g. the coverage areas of nodes are disjointed, then

 F m~ , , ( ˜Y ™— / } Y “( /; o sY — G }

on segment for bandwidth extension. To meet node ’s
bandwidth requirement for transmission, we have
2 Note
that different scheduling policies can be used for message trans-
mission and reception. .
 Ctf
D CgfuD m }šDD kjCEf‘DrD vD k‚CgffWD š} D kjf;D D k-D
Proof: Let be the tour generated by the TSP al- 2000
UN, uniform traffic
UN, non-uniform traffic
gorithm in the MFR Algorithm and be an optimal TSP CN, uniform traffic
CN, non-uniform traffic

Y†› m Y
tour. Then we have . From the description of

›
1500
the MFR Algorithm, we have
where is the amount of tour extension to meet all nodes’

Average Delay
bandwidth requirements. 1000

in [3]. Let *
We now prove our claim using the technique proposed
be the area swept by a disk of radius
k k ~
 ™  m * *  meœ XT Do k-T D Y œ ™   k
whose center moves on route . Since route visits each 500

node’s coverage area, includes the coverage area of each

mž Ÿ Y œ
node. We have where is the
0
number of nodes. Thus, . Note that may 20 40 60 80 100 120


Number of nodes
not visit each node’s position since the ferry can transmit
to nodes within distance . A TSP tour which visits each

 k Figure 3. Message delay vs. network size

D k-DD k‚Y f‘~ D ~ a ,;m ( ,;( ŸY¢/Ÿ¡ D k-/D D k-D Y  —  k f

node’s position can be obtained from route by making a
detour of length of at most to visit each node’s position.

}šD k‚f£D Y D k-D mm , , ( YsY¤¡ Ÿ¡ / } Y Y (“— /D k-D Y — N}

This tour is of length . Since 16000
UN, 100 nodes

D CED m
CN, 100 nodes
is an optimal TSP tour, . Therefore, 14000
UN, 80 nodes
CN, 80 nodes

GF m T FU T k
UN, 20 nodes
we have . CN, 20 nodes
12000

TJUo T  o y
By Property 2, . For route , since the average

B¥
10000

TJUo T GF mn~ ,,;( Y Ÿ¡ /;} Y ( /;B¥ o Y ¡ U

waiting time of messages in the sending node is ,
Average Delay

8000
. So we prove .

& ,,( Ÿ /}
When the number of nodes is large, will be close to U¡ ` 6000

~ Y˜¡ Y (“/
. Thus the MFR Algorithm has an approximation ratio of 4000

for large networks with disjointed nodes. 2000

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
4 Preliminary Experimental Results Network Load

Figure 4. Message delay vs. network load

In this section we present preliminary experimental re-


sults on the performance of the MF scheme in the stationary

œ &&&ƒ¦ œ &&&ƒ
node case. We model nodes distributed in a rectangular

~ &ƒ§wA (&N&rƒ
area in dimension. A single ferry is de- We also experiment with different traffic models in the
ployed which moves at a speed of . The transmission simulations. In the uniform traffic model, the traffic be-
range of both the nodes and the ferry is unless oth-
O« ‘¬ ª 
tween any two nodes has equal bandwidth which is set to
&’8(
«O¬­ 
erwise noted and the transmission rate is 10Mbps. We sim- Kbps. In the non-uniform traffic model, flows

«O‘¬ ª 
ulate a fair sharing scheduling policy in which each node are sending at rate Mbps and the rest are sending at
gets equal access to the ferry when the ferry is within range Kbps. This non-uniform traffic model represents the
of multiple nodes. Two kinds of node distributions are con- case where some flows transmit much more data than oth-
sidered in the experiments. ers. In both traffic models flows are generated at constant
 bit rates.

‹ ,W&’ &¨ N& &N/

Random Uniform Node Distribution (UN). In this dis- Given node locations and the traffic models, we use the

œ
tribution, each node’s and coordinates are chosen MFR Algorithm to compute the ferry route. In our im-
randomly from interval . plementation of the MFR Algorithm, we choose the near-
 est neighbor algorithm with 2H-opt to compute the starting

case, we first compute 9w ~ &

Random Clustered Node Distribution (CN). In this
cluster centers whose co-
,W&’ œ &&N&N/
TSP tour and then use local optimization technique 2H-opt-
delay to reduce the message delay.


ordinates are chosen randomly from interval
For each of the nodes, we calculate its position by
. We evaluate the performance of the MF scheme under
different network sizes and traffic loads. The main met-

œ & &&OwN© 
computing two normally distributed variables, each ric used in our evaluations is average message delay. Fig-
multiplied by , and adding them to the co- ure 3 shows the average message delay with different net-
ordinates of a randomly chosed cluster center. work sizes. In general, as the number of nodes increases,
so does the message delay. And the clustered node distri- [4] Z. Haas and M. Pearlman. The performance of query control
butions tend to have smaller delay as compared to the uni- schemes for the zone routing protocol. In SIGCOMM, 1998.
form node distributions. This is because clustering of nodes [5] D. Johnson and D. Maltz. Dynamic source routing in ad-hoc
tends to shorten the ferry route which results in smaller wireless networks. In SIGCOMM, August 1996.
[6] D. Johnson and L. McGeoch. Experimental analysis of
message delay. The important point is that, in the scenar-
heuristics for the stsp. The Traveling Salesman Problem and
ios we study here, the message ferry routing can achieve
its Variations, 2002.

œ&
reasonable performance even when the nodes are extremely [7] Y. Ko and N. Vaidya. Location-aided routing (lar) in mobile

œ (“&N&&¯°&rƒ®>£±X A œ &N&&ƒ
disconnected. For example, with nodes randomly dis- ad hoc networks. In MOBICOM, November 1998.
tributed in a area, each node can send mes- [8] Q. Li and D. Rus. Sending messages to mobile users in dis-
sages at rate with about 1071 second delay. With- connected ad-hoc wireless networks. In MOBICOM, August
out any communication infrastructure, we believe this delay 2000.
is inherent given the dimension of the deployment area and [9] G. Nemhauser, A. R. Kan, and M. T. (editors). Optimization.
the limited transmission range. 1989.
[10] C. Papadimitriou. The euclidean traveling salesman problem
Figure 4 shows the effect of network load on the mes-
is np-complete. Theoret. Comput. Sci., 4:237–244, 1977.
sage delay. We define network load as the ratio between [11] V. Park and M. Corson. A highly adaptive distributed al-
total transmission bandwidth generated by nodes and the gorithm for mobile wireless networks. In INFOCOM, April
transmission rate of the wireless interface. The average 1997.

, œ &O² ´³ &O²E/ (

message delay increases as the network load increases es-
pecially when the network load is close to . However, for
modest network load , the increase of delay
[12] C. Perkins and P. Bhagwat. Highly dynamic destination-
sequenced distance-vector routing (dsdv) for mobile com-
puters. Computer Communication Review, 24, October
is very slow suggesting that nodes can send at higher rates 1994.
without significant delay penalty. We also evaluate the per- [13] C. Perkins and E. Royer. Ad hoc on-demand distance vec-
tor routing. In 2nd IEEE Workshop on Mobile Computing
formance of the MF scheme under different transmission
Systems and Applications, February 1999.
ranges which show similar results.
[14] R. Sivakumar, P. Sinha, and V. Bharghavan. Cedar: a core-
extraction distributed ad hoc routing algorithm. IEEE Jour-
5 Conclusion nal on Selected Areas in Communications, 17:1454–1465,
August 1999.
In this paper we presented the Message Ferrying scheme [15] A. Vahdat and D. Becker. Epidemic routing for partially-
to solve the data delivery problem in highly-partitioned connected ad hoc networks. Technical report, Duke Univer-
sity, 2000.
wireless ad hoc networks. It is a proactive scheme in
which a set of nodes called message ferries take the re-
sponsibility of carrying messages and providing connec-
tivity for all nodes in the network. By introducing non-
randomness in nodes’ proactive movement and exploiting
such non-randomness to deliver messages, the MF scheme
provides regular connectivity in a disconnected network and
improves data delivery performance without global knowl-
edge of each node’s location. In this paper, we also study
the ferry route design problem for the case where nodes are
stationary. We show that this problem is NP-hard and pro-
vide an efficient algorithm to compute the ferry route.

References
[1] J. Bentley. Fast algorithms for geometric traveling salesman
problems. ORSA Journal on Computing, 4:387–411, 1992.
[2] J. Davis, A. Fagg, and B. Levine. Wearable computers as
packet transport mechanisms in highly-partitioned ad-hoc
networks. In International Symposium on Wearable Com-
puting, October 2001.
[3] A. Dumitrescu and J. Mitchell. Approximation algorithms
for tsp with neighborhoods in the plane. In Twelfth ACM-
SIAM Symposium on Discrete Algorithms, pages 38–47, Jan-
uary 2001.