ROUTING PROTOCOLS FOR DELAY TOLERANT NETWORKS: SURVEY AND PERFORMANCE EVALUATION

International Journal of Wireless & Mobile Networks (IJWMN) Vol. 7, No. 3, June 2015
DOI : 10.5121/ijwmn.2015.7305 55
ROUTING PROTOCOLS FOR DELAY TOLERANT
NETWORKS: SURVEY AND PERFORMANCE
EVALUATION
Luming Wan, Feiyang Liu, Yawen Chen, and Haibo Zhang
Department of Computer Science, University of Otago, Dunedin, New Zealand
ABSTRACT
Delay Tolerant Networking (DTN) is a promising technology that aims to provide efficient communication
between devices in a network with no guaranteed continuous connectivity. Most existing routing schemes
for DTNs exploit the advantage of message replication to achieve high message delivery rate. However,
these schemes commonly suffer from large communication overhead due to the lack of efficient mechanisms
to control message replication. In this paper we give a brief survey on routing protocols designed for
DTNs, and evaluate the performance of several representative routing protocols including Epidemic, Spray
and Wait, PRoPHET, and 3R through extensive trace-driven simulations. Another objective of this work is
to evaluate the security strength of different routing schemes under common DTN attacks such as the black
hole attack. The results and analysis presented in this paper can provide useful guidance on the design and
selection of routing protocols for given delay-tolerant applications.
KEYWORDS
Delay Tolerant Network, Routing, Performance Evaluation, Communication Overhead.
1. INTRODUCTION
The increasing popularity of ubiquitous computing and communication has led to a huge desire
on data exchange between wireless mobile devices, e.g. cell phones, laptops, tablets, and other
portable devices, regardless of whether any guaranteed end-to-end connection exists. Delay
tolerant networking (DTN) addresses the technical challenges on communication between devices
that lose continuous connectivity due to mobility. Currently, DTNs have been applied to a vast of
areas, including vehicular networks [1], wildlife tracking [2], and social network analysis [3], etc.
However, DTNs are still appearing with numerous of limitations. For example, due to the lack of
synchronous end-to-end connectivity, mobile devices have to carry the messages and forward
them opportunistically upon encountering the destinations, or forward messages to other relays to
help the delivery. This could result in incredible long transmission delay and low message
delivery rate. In addition, the constraints on wireless devices, such as storage capacity,
communication bandwidth, and battery power, can significantly impact the successful message
delivery rate. Hence, routing protocols for DTNs should be able to adapt to the network variation
and be efficient to make use of the available hardware resources.
Recent studies exhibit that most of existing routing protocols designed for DTNs appear with high
similarity in concepts, but their performances are significantly disparate. Epidemic [4] performs
robustly as it blindly floods messages to all devices in the network. Spray and Wait [5] is a
flooding-controlled version of Epidemic, as it simply limits the amount of messages that can be

56
flooded to the network. In most prediction-based schemes, such as PRoPHET [6], MaxProp [7]
and PER [8], the message forwarding decisions are made based on a quality metric called
encounter predictability which is accumulated from past encounters. However, the above
prediction-based schemes ignore some important encounter information, such as the contact time
and contact duration. Some studies demonstrate that human activities appear with high repetition,
such as weekly meetings. The performance of routing schemes could be enhanced significantly
through the exploitation and utilization of those regular patterns. 3R [9] is a fine-grained history-
based routing scheme, by which the contact time of pair-wise encounters are perfectly recorded.
Due to the maintenance of fine-grained encounter history, the encounter predictability in 3R is
time-dependent, that is, it can be calculated based on only the past contacts that occurred in the
same period as the lifetime of the packet, rather than making a long-term average estimation as in
most prediction-based protocols such as PRoPHET. In addition, 3R is a forwarding-based scheme.
Each node always forwards the original message to the next good relay rather than transmitting
replicas as in all the protocols mentioned above. However, single-copy for each message is not
ideal for DTNs since message could be easily lost due to unstable connectivity, message buffer
overflow, and other uncertainties.
It is difficult to design an absolutely perfect routing protocol that suits to all DTN applications.
The performance of a routing protocol can be affected by a large number of factors, such as the
popularity and active rate of nodes, different setting of parameters for mobile devices, etc. In this
paper, we evaluate the performance of several most popular routing protocols, including
Epidemic, Spray and Wait, PRoPHET, and 3R through trace-driven simulations. Also, we
investigate the impact of different parameters of Spray and Wait and PRoPHET to achieve their
best performance. The aim of this work is to provide some important guidance on routing
protocol design and selection for delay tolerant networks.
On the other hand, it is challenging to develop a secure routing scheme for DTNs due to the vast
constraints such as hardware limitation and unstable end-to-end connectivity. Oversimplified
routing protocol are highly vulnerable to be attacked. Byzantine attacks[14] summarized several
typical categories of attacks in DTNs. Most existing routing protocol for DTN cannot address the
Byzantine attacks. In this paper, we also evaluate the security strength of several protocols with
the black hole attack, which is a representative class of Byzantine attacks.
The reminder of this paper is organized as follows: Section 2 introduces the protocols we are
going to evaluate in detail. Section 3 introduces the data traces and experiment setup. Section 4
shows the performance evaluation in two perspectives. Firstly, we present the impact of different
parameter settings for Spray and Wait, and PRoPHET protocols. Then, we evaluate the
performance of all the protocols with their optimal parameter settings. This includes the general
performance comparison, the average message delivery latency of each protocol, and the
performance difference of the protocols under various packet longevity. In Section 5, we measure
the security strength of the protocols. The paper is concluded in Section 6.
2. ROUTING PROTOCOLS
This section introduces the details of four protocols, Epidemic, PRoPHET, Spray and Wait, and
3R, that will be evaluated in our simulations.

57
2.1. Epidemic
Epidemic [4] is a pure flooding-based routing protocol for DTNs. Each device (source or relaying
node) always propagates message replicas to all of the contactable nodes until the message is
received by the destination node or the message deadline expires. Due to the nature of unlimited
flooding, Epidemic is able to achieve the best message delivery rate when each node has an
infinite memory buffer, but its performance deteriorates significantly when each node only has
limited resources. This is because numerous of message replicas have to be dumped due to
memory overflow.
2.2. Spray and Wait
Similar to Epidemic, Spray and Wait [5] is also a flooding-based routing protocol, but it controls
the message flooding by directly limiting the number of replicas that an original message can
produce and forward. In spray phase, each message can only propagate a constant number L of
replicas. The source node forwards only one replica to the next node it encounters. The source
node keeps propagating replicas until there is only one left, then the source node and all relays it
encountered before switch to the wait phase. That is, those messages only wait for the destination
node encountering their holders rather than further spreading replicas to other relays.
There is a derivation version of Spray and Wait called Binary Spray and Wait. Instead of perform
pure flooding like Spray and Wait, Binary Spray and Wait propagate its message copies in a
binary tree manner. Suppose a message is allowed to spread L replicas, the source node will
generate L replicas in the beginning and spread them among encountered relays. Suppose a node
A has n (1 < n < L) replicas of message m, and it encounters a node B with no replica of m
previously. A will send 2
n
   replicas to B and keep the rest of 2
n
   replicas to wait for more
relaying nodes. When only one replica is left in A or any other relaying node, they stop spreading
any replica and wait for encountering the destination node. Spray and Wait can achieve a trade-
off between the delivery rate and transmission overhead by setting a proper limitation of L.
2.3. PRoPHET
PRoPHET [10] is a prediction-based scheme, and it is one of the few DTN routing protocols that
have an IETF draft. It implements a quality metric called encounter predictability to measure the
capability of the encountering nodes whether or not it can transmit the message to the destination.
A replica of a message is propagated to the encountering node if it has a higher predictability than
its holder. This guarantees that each time a message can always be propagated to a better relay.
The encounter predictability is integrated in three perspectives, which are direct probability,
transitivity and ageing. Direct probability is the probability of transmitting a message directly
between two nodes. It is updated whenever two nodes directly encounter each other as follows:
ܲሺܽ, ܾሻ = ܲ௢௟ௗሺܽ, ܾሻ + ൫1 − ߜ − ܲ௢௟ௗሺܽ, ܾሻ൯ × ܲ௘௡௖௢௨௡௧௘௥ , (1)
where Pold(a,b) is the encounter probability of node a and b before the current encounter occurs,
Pencounter ∈ [0,1] is a scaling factor at which the probability increases on encounters, and δ is a
small positive value to set an upper bound on P(a,b).
Transitivity estimates the probability of indirect contact that through multiple-hop relays, as
shown below:

58
ܲሺܽ, ܿሻ = ܲ௢௟ௗሺܽ, ܿሻ + ൫1 − ܲ௢௟ௗሺܽ, ܿሻ൯ × ܲሺܽ, ܾሻ × ܲሺܾ, ܿሻ × ߚ , (2)
where P(a,c) is the encounter probability between node a and c through relay node b, Pold(a,c) is
the probability from previous time of update, and β ∈ [0,1] is a scaling factor that decides how
large the impact of transitivity should have on the encounter predictability.
PRoPHET uses an ageing mechanism to decay the encounter predictability and eliminate the
long-time inactive nodes, as shown below:
ܲሺܽ, ܾሻ = ܲ௢௟ௗሺܽ, ܾሻ × ߛ௞
, (3)
where ߛ ∈ [0,1] is the ageing constant, and K is the number of time units elapsed since the last
time P(a,b) was aged.
As shown in the above Equations, there are three important parameters for PRoPHET protocol:
scaling factors for direct delivery Pencounter, transitivity β, and the ageing constant ߛ. They have
significant influence on the performance and overheads of PRoPHET scheme. PRoPHET can be
an efficient routing protocol by setting the proper parameters if the encountering patterns in
DTNs are predictable, because the messages and replicas are only forwarded to the relays with
more benefits on message delivering. At present, there are two versions of PRoPHET protocol:
PRoPHET'03 [6] and PRoPHET'12 [10]. They have different recommendation settings for these
parameters, while the principle of message forwarding for both versions are basically identical.
PRoPHET'12 [10] has slight improvement on its routing mechanism by filtering out the transient
contacts. These transient contacts are not long enough for nodes to exchange messages, and thus
they should not be counted into history and take effect to the predictability.
2.4. 3R
3R [9] characterizes the encounter history in a fine-grained form by storing more information for
each encounter, including the nodes that the encounter occurs, the start time and the end time of
each encounter. Each node maintains a table for the encounter information of its contacted nodes.
The fine-grained history in a node will group the contacts according to different types of the day
{weekday, weekend}. Each item in the table represents a fixed time interval of the real world (e.g.
1 hour), and it records the overall contact frequency between the table holder and the specified
node at exact the same period in the past. Suppose a message m has a lifetime that spans k slots.
Let Fa
i be the number of contacts of node a that occurred at a time slot i in the past, and Fi
ab be the
number of contacts occurred between node a and b at time slot i, the estimated contact probability
pi
ab at slot i is:
‫݌‬௔௕
௜
=
ிೌ್
೔
ிೌ
೔ . (4)
Suppose the lifetime of a message m spans k time slots, the overall probability that message m can
be delivered from node a to node b before it expires is estimated by integrating the probabilities
of these k slots as follows:
ܲ௔௕ሺ݇ሻ = 1 − ∏ ሺ1 − ‫݌‬௔௕
௜
ሻ௞
௜ୀ଴ . (5)
Due to the maintenance of fine-grained history, a time-dependent forwarding prediction is
enabled by only figuring out the delivery probability within the lifetime of a message, rather than

59
calculating a long-term average estimation. Similar to PRoPHET, 3R also compares the
predictability and forwards messages to nodes with higher delivery probability. Whereas, 3R only
directly forwards the original message to the next relay, instead of sending a replica as in
PRoPHET. Due to the predictability calculation based on fine-grained contact history and single-
copy message forwarding, 3R can effectively reduce delivery overheads with no much loss on
message delivery rate.
3. Data Traces and Experiment Setup
3.1. Data Traces
We evaluate the above four representative routing protocols for DTNs with extensive trace-driven
simulations. To increase the accuracy and reliability of our evaluation, we use two realistic DTN
data traces: INFOCOM [11] and MIT Reality [12], which are obtained from the open-source
website CRAWDAD [13].
INFOCOM: This data trace consists of Bluetooth sightings and it records 4 days of contacts
occurred during the INFOCOM'05 conference. In the experiment, 24 internal devices
monitored the network and more than 200 external devices were discovered. Each device has
a scan granularity of 120 seconds, and each scan lasted for 5 seconds.
MIT Reality: This trace contains the contact information including communication,
proximity, and location from faculties and students at MIT over the course of the 2004-2005
academic year. To collect this trace, 89 devices were used to record the contacts, and more
than 20,000 devices were discovered. Each device scans to discover neighbors every 5
minutes. In our simulations, we use the devicespan subtrace that records Bluetooth contacts
for 1 month.
3.2. Experiment Setup
In our simulations, the routing schemes are evaluated using the same message trace and contact
trace. To avoid the negligible delivery rate caused by numerous long-time inactive nodes which
actually seldom communicate, we randomly select the source node and the destination node of a
message from 20 most active nodes during the lifetime of this message. In the INFOCOM trace,
each source generates a message with the probability of 0.15 in every 600 seconds, whereas the
probability is 0.6 in the MIT Reality trace. The lifetime of each message varies from 2 hours to 3
days if not specially notified. The size of each generated message is randomized from 2k bytes to
100k bytes.
For the MIT Reality trace, the network is warmed up for one week at the beginning of each
simulation run. This avoids the inaccurate prediction on encounter probability for prediction-
based protocols, i.e. PRoPHET and 3R, when the network is just booted up. Also, we reserve 3
days at the end of simulations to avoid any message left in the network after the simulation is
completed. During these warm-up and sinking periods, messages are not allowed to be generated.
This process is not available for the INFOCOM data trace, since it only contains 4 days of contact
traces.
4. PERFORMANCE EVALUATION
In this section, we firstly show the performance of Spray and Wait and PRoPHET with different
parameter settings using the INFOCOM data trace, since they have some specific parameters.

60
Then we investigate the performance of all four protocols, i.e. Epidemic, Spray and Wait,
PRoPHET, and 3R, through trace-driven simulations. The performance of these four protocols are
investigated in three perspectives: different memory buffer sizes, average message delivery
latency, and packet lifetimes. The performance is evaluated using the following three metrics:
average message delivery rate: the proportion between the number of successfully delivered
messages against the number of original messages.
message overhead (MO): the ratio of the number of replicas against the number of original
messages.
communication overhead (CO): the ratio of the difference between the number of delivered
messages and hop-by-hop transmissions to the number of delivered messages.
The parameter setting for Spray and Wait is slightly different in each data trace in accordance
with the characteristic of the data trace. The Binary Spray and Wait protocol is implemented in
the entire experiments as it performs slightly better than its original version. Due to the huge
difference on popularity of nodes in these two traces, the maximum number of replicas for Binary
Spray and Wait with INFOCOM trace is set to 5, and 70 with MIT Reality trace. The parameters
in PRoPHET are configured with the values provided in PRoPHET'12 [10] by default. All the
nodes are set with infinite memory buffer if not specified. All the protocols dump messages using
the First In First Out policy when the message buffer is full, that is, the first message in the buffer
queue (oldest) is dumped to prevent buffer overflow.
4.1. Impact of Parameters for Spray-and-Wait and PRoPHET
Spray-and-Wait: Fig. 1 shows the performance of the original and binary version of Spray and
Wait routing protocol with different maximum number of replicas. In Fig. 1(a), it can be seen that
the original Spray and Wait achieves higher delivery rate than the binary version when the
maximum number of replicas is less than 10. With the further increase on the number of replicas,
the binary version performs better than the original version. That is because the original scheme
can only spread the replicas around the neighbours, while the binary scheme can spread the
replicas more widely in a binary-tree manner if enough number of replicas are allowed. We can
see the delivery rate of both protocols remain steady when the maximum replicas exceeds 40.
The highest average delivery rate achieved by the original and the binary versions are 0.9314 and
0.9382, respectively. That means the number of replicas are enough for two Spray and Wait
protocols with no further benefits for more replicas. Meanwhile, MO and CO for both schemes
appear with the similar increasing trend as shown in Fig. 1(b) and (c). The MO and CO of the
original Spray and Wait increase apparently with the maximum replicas less than 40, and they
stay constant afterwards also due to the limitation on the number of neighbours, whereas both two
kinds of overheads of Binary Spray and Wait grow almost linearly since the replicas can be
spread to more relay nodes on the network. From these figures, it can be seen that with a large
number of replicas, the multi-hop routing of Binary Spray and Wait is able to perform better than
the only two-hop routing in original version, but it suffers almost twice of overheads than the
original version.

61
Fig. 1. The performance of Spray and Wait with different number of message replicas, (a) delivery rate, (b)
message overhead, (c) communication overhead.
PRoPHET: The performance of PRoPHET routing protocol is critically impacted by the
parameters shown in Table 1. PRoPHET provides totally different recommended parameters in
their 2012 version [10] compared to the 2003 version [6].
Table 1. PROPHET recommended parameters
Parameters Pencounter Β Γ
PRoPHET 2003 0.75 0.25 0.98
PRoPHET 2010 0.5 0.9 0.999
Fig. 2 shows the performance of PRoPHET'03 and PROPHET'12.
In these experiments, each node has infinite memory buffer to store messages and replicas. The
principle of message forwarding for these two versions are basically identical. It can be seen that
in Fig. 2(a), the delivery rate is increased by 0.5% with the parameters of PRoPHET'12 in
comparison with PRoPHET'03. A larger ageing parameter γ makes PROPHET'12 better to
tolerate the variation of contact patterns and achieve higher delivery rate. However, both MO and
CO are raised by 6.1% and 5.4% respectively as shown in Fig. 2(b). This is because a larger β in
PRoPHET'12 leads to higher impact on the transitive connectivity, and a lower Pencounter reduces
the impact of direct delivery. Hence, more replicas need to be generated and forwarded by the
relay nodes.
4.2. Performance Evaluation with Different Memory Sizes
Fig. 2. PRoPHET'03 versus PRoPHET'12, (a) delivery rate, (b) overheads.

62
In each mobile device, the memory buffer used for message delivery always has a limited size.
The available memory has huge influence on the delivery rate and overheads. We evaluate the
performance of four routing protocols with different memory sizes using INFOCOM and MIT
Reality traces.
(1) INFOCOM Trace: In Fig. 3(a), we evaluate the influence of memory size on the delivery
rate for different routings. It can be seen that the delivery rates of all four routing protocols have a
sharp increasing trend with a small memory buffer (<10MB), and then they tend to remain stable
even if the memory buffer size keeps growing. Binary Spray and Wait has more outstanding
performance than other three routings when each node has a memory buffer of 80M bytes or less.
It reaches its maximum delivery rate of 0.923 with the memory buffer size increasing to 60M
bytes, and the delivery rate tends to be stable afterwards. That is because it is a controlled
flooding scheme (5 replicas) and has less message dropped due to memory overflow than other
routings. When the memory size exceeds 60MB, there will be no message dumps for this routing
and the delivery rate will be stable. Similarly, 3R also has better performance with a smaller
memory buffer size compared with Epidemic and PRoPHET. It has less requirement on memory
size because of single-message and its limitation on message forwarding. However, it reaches the
maximum delivery rate of 0.792 at the memory size of 40M bytes, which is much smaller
compared with the other three protocols. That is because 3R only forwards the original message.
Although the delivery rates of Epidemic and PRoPHET raise slowly, their maximum delivery
rates are higher than 3R and Spray and Wait (0.941 and 0.930 respectively) since they are
memory-hungry schemes. Epidemic protocol needs large amounts of memory for unlimited
flooding and can achieve the highest delivery rate with enough memory. PRoPHET also needs a
large memory size for transmitting the amounts of messages and replicas.
Fig. 3(b) and (c) evaluate the overheads with different memory sizes. We can see that Epidemic
always has the highest MO and CO due to the nature of unlimited flooding. Since it needs to
generate more replicas when encountering relay nodes with high predictability, PRoPHET also
has higher overheads, which is almost half of the overheads as Epidemic. Because the maximum
number of replicas in Binary Spray and Wait is set to 5, the maximum MO and CO of Binary
Spray and Wait scheme tend to be 5 even with a much larger memory size. Since 3R is a single-
message forwarding-only scheme, there is no message overhead. The CO of 3R is also lowest
because the message is only forwarded to a node with higher predictability according to fine-
grained encounter history. The CO of Binary Spray and Wait and 3R appears with a decreasing
trend as the memory size increases, since they have limitations on message forwarding, and a
larger memory size increases the number of delivered messages by reducing message dropping.
Fig. 3. Performance evaluation with different memory sizes using INFOCOM trace, (a) delivery rate, (b)

63
Fig. 4. Performance evaluation with different memory sizes using MIT Reality trace, (a) delivery rate, (b)
(2) MIT Reality Trace: As shown in Fig. 4(a), the delivery rate of each routing protocol has
similar trend as the memory size increases. However, PRoPHET and 3R have better performance
than Epidemic and Spray and Wait when the memory size is small (<5MB). Since the MIT
Reality trace lasts for 1 month and these two routings can make use of the contact history to get
more accurate prediction for message forwarding, and a smaller memory is relatively enough due
to less replicas and message forwarding. As the memory size increases, Epidemic and Spray and
Wait achieve higher delivery rate because a large number of flooding replicas can be tolerated in
the memory with few of message dumps.
Similarly, the message overhead (MO) in Fig. 4(b) and communication overhead in Fig. 4(c) vary
in the same way as the memory size increases. Since there are more nodes in the MIT Reality
data trace than that in the INFOCOM trace, the overheads of Epidemic are significantly increased.
That means Epidemic is not suitable for DTNs with a large number of devices. We can see that
the overheads of Spray and Wait increase with the increase of maximum allowed replicas (from 5
in INFOCOM trace to 70 in MIT Reality trace). PRoPHET and 3R protocols have lower overhead
for both data traces, which is determined by the popularity of nodes and the activeness of each
node.
From this set of simulations, it can be seen that flooding-based routing protocols can achieve
better delivery rate, but require much larger message memory. Whereas, the prediction-based
routing protocols are more memory efficient, and can achieve desirable delivery rate with very
low overheads if they have accurate prediction on the contact patterns.
4.3. Average Latency
We also conduct a set of simulations to measure the routing efficiency of each DTN protocol in
term of the average message delivery delay. The message delivery latency is defined as the time
span from the time the message is generated until the time when the original message or one of its
copy reaches the destination. We evaluate the average message delivery latency of the four
routing schemes using the INFOCOM and MIT data traces separately with different memory
capacities.

64
Fig. 5. Average message delivery latency with different memory capacity using trace (a) INFOCOM, (b)
MIT Reality
(1)INFOCOM trace: In Fig.5 (a), all of the protocols appear with a similar increasing trend.
Epidemic spends the least average time on message delivery for all cases studied because it takes
the advantage of blind flooding. Under the situation of a large memory buffer at each node,
Epidemic has the shortest delay because messages can survive longer at nodes. Large number of
message copies of a single message produce a higher chance for the message to reach its
destination through opportunistic encountering between nodes, rather than waiting for frequent
contact nodes like PRoPHET and 3R. On the other hand, when nodes have very limited memory
buffer, most of the message copies have to be discarded because of memory overflow. Thus the
delivery rate is incredible low for Epidemic. However, these limited amount of deliverable
messages still takes the benefit of flooding. The messages can be flooded to the destination
through a relative short path, so that Epidemic still has the shortest delay with the extremely
limited memory at nodes. PRoPHET costs slightly more than Epidemic. It takes the second place
because of its accurate routing prediction. The performance of 3R is relatively poor because of its
single-message-forward mechanism. The use of only a single message is hard to find out the
shortest path to the destination in DTN.
(2)MIT Reality trace: Similar to Fig.5 (a), Epidemic and 3R still get the first and the last place
on the average message delivery delay. However, Spray and Wait outperforms PRoPHET in MIT
Reality trace. This is because this trace contains a much larger node population. The upper limit
of message copy is set to 70, which is a huge increase against 5 in INFOCOM trace. Therefore,
Spray and Wait performs almost the same degree of flooding as Epidemic, so that messages can
be rapidly delivered to their destination through such flooding.
From the results above, flooding-based routing protocol always gain a lower latency than
prediction-based all the time. However, combining the result in Fig. 3 and Fig. 4, flooding-based
protocols have a poor performance when nodes have limited buffer. In contrast, flooding-based
protocols obtains incredible high overheads with large memory buffer. It is necessary to
comprehensively inspect the routing protocols to make a fair comparison, rather than only
observe a single aspect.
4.4. Performance Evaluation with Different Packet Lifetimes
Each message may experience an unpredictable delivery latency in DTNs due to the lack of
guaranteed continuous end-to-end connectivity. Therefore, in order to achieve more effective
communication for DTNs, each message needs a suitable lifetime either by users setting or

65
default configuration of specific devices. In this set of simulations, we evaluate the influence of
different packet lifetimes for the routing algorithms using INFOCOM and MIT Reality data
traces respectively. We assume each node has an infinite buffer size.
(1) INFOCOM Trace: Fig. 6 shows the delivery rate and overheads of four routing protocols
with different setting of packet lifetimes using the INFOCOM trace. In Fig. 6(a), it can be seen
that for each routing protocol the delivery rate appears with a dramatic increase when the packet
lifetime increases from 6 to 30 hours. For example, the delivery rate increases from 0.454 to
0.832 for 3R routing. The reason is that the probability of directly encountering the destination or
the message delivered to the destination by multi-hop forwarding grows when the packet lifetime
increases. This increasing trend slows down when the packet lifetime exceeds 30 hours. That is
because the data trace only lasts 4 days, and the contact pattern is short-time dependent. The
delivery rate will not increase even with a large packet lifetime since the destination node may
only be active in a short time period. The overheads shown in Fig. 6(b) and (c) also demonstrate
that, for Epidemic and PRoPHET, the overheads increase as the packet lifetime increases, since
more replicas can be generated and forwarded to the destination. Whereas, the lifetime is enough
long for message delivery when it exceeds 30 hours in this trace. The overheads of Spray and
Wait and 3R keep stable with different packet lifetime, because the number of replicas is limited
in Spray and Wait and 3R is a single-message forwarding based scheme.
Fig. 6. Performance evaluation with different packet lifetimes using INFOCOM trace, (a) delivery rate, (b)
Fig. 7. Performance evaluation with different packet lifetimes using MIT Reality trace, (a) delivery rate, (b)

66
(2) MIT Reality Trace: The simulation results with MIT Reality trace are similar as shown in
Fig. 7. It can be seen in Fig.7 (a) that the delivery rate for all the routing protocols keep growing
as the packet lifetime increases. It is because MIT Reality trace lasts for 1 month and it is long-
time dependent, thus a longer packet lifetime always provides a larger probability of successful
delivery. In this simulation, as shown in Fig.7 (b) and (c) the overheads of Epidemic increase
dramatically both due to longer packet lifetime and larger device population. Whereas, the
overheads variations for other routing protocols are similar to the results using INFOCOM data
trace.
This set of simulations demonstrate that a longer packet lifetime can provide a higher delivery
rate for DTNs, since there is more chance to encounter the destination node either by direct
contact or multi-hop message forwarding. The overheads of flooding-based routings increase
significantly, while the prediction-based routings are more efficient, especially when the DTNs
have long-time dependent contact patterns.
5. SECURITY EXPERIMENTS
5.1. Black Hole Attack and Attack model
We also investigate the security issue inside DTN protocols. Due to the simplicity of routing
design, the protocols we evaluated appear with bare capability to handle hostile attacks from the
unstable DTN environment. To test their security strengths, we perform black hole attack which
interrupt message spread in the network. Similar to the black hole in universe, the black hole
nodes in DTN assimilate all incoming messages and never forward them to others. In the
simulation, we randomly select a fixed amount of attacker nodes from the entire population. The
amount of attacker nodes is decided by a coefficient called attack opportunity, which measures
the badness of DTN environment. The attacker nodes do not originate any message, and all
packets will be destroyed after received by these attacker nodes. To make a fair comparison for
history-based protocols, the attacker nodes still store and update their contact history, thus
sometimes the intruders might be treated as a good relay for some messages. To remove the
impact of further losing packets such as message discard from memory overflow, all of the nodes
in the simulation have infinite buffer size.
5.2. Security Test with Black Hole Attack
(1)INFOCOM: Fig. 8 evaluates the performance of protocols with a steadily increased attack
opportunity using the INFOCOM trace. From Fig. 8 (a), we can see that all of the protocols have
very similar decrease trend when attack opportunity grows, but 3R performs much worse than
others. Due to the nature of single-message-forward scheme, the lost messages are always
irretrievable after they are consumed by attacker nodes. This brings a huge performance degrade
for 3R. The rest of the protocols gain a higher and very similar performance. This is because they
are all replication-based schemes. Destroying a message replica by an attacker node is not a
critical problem for them since the destroyed message might have more copies exist in
somewhere else in the network. In (b) and (c), Epidemic still has the highest overheads because
of its blind flooding. There is a very interesting point is 3R and Spray and Wait which have
obvious increase on communication overhead at the very late stage. The reason of this for 3R is
that nodes forward original messages through multiple hops and most of them eventually meet
attackers, but unfortunately the transmission history of the destroyed messages still have to be
counted into the overheads. This result in 3R has a very poor delivery rate but still gains a great
grow on communication overhead. The reason for Spray and Wait is that it propagates message

67
copies as a binary tree. The attacker nodes might receive a large number of copies for a single
message (e.g. a attacker nodes could receive 2/n of copies at most when it directly receives them
from the source node) and then destroy all of them. Hence, Spray and Wait is very sensitive to the
amount of message copies. As the attack opportunity grows, the number of successfully delivered
message drops faster than the number of message copies it generated. Hence, the CO of Spray
and Wait is proportional to attack opportunity.
Fig. 8. Performance evaluation with different black hole attack opportunity using INFOCOM trace, (a)
delivery rate, (b) message overhead, (c) communication overhead.
Fig. 9. Performance evaluation with different black hole attack opportunity using MIT Reality trace, (a)
delivery rate, (b) message overhead, (c) communication overhead.
(2)MIT Reality: Fig. 9 shows the result using the MIT Reality trace. It can be seen that the
performance differences between protocols appear much more obvious than that in the
INFOCOM trace. This is because MIT trace contains larger amount of noise and inactive nodes.
In Fig. 9 (a), 3R still has an unexpected performance as above. PRoPHET performs slightly worse
than Epidemic and Spray and Wait when the attack opportunity is less than 0.4, but overtakes
them afterward. It is because PRoPHET takes the advantage of accurate routing prediction
compared to flooding-based protocols. Epidemic has a little bit higher delivery rate than Spray
and wait with a incredible high attack opportunity because Epidemic do not have message
replication limit. In Fig. 8 (b) and (c), it can be seen that Epidemic still has very high overheads.
Different from previous experiments, PRoPHET and Spray and wait have very close overheads.
This is because the node population in MIT is much larger than that in INFOCOM, and Spray and
Wait needs more message copies to achieve a relative high delivery rate.
In this experiments, it can be seen that single-message-forward protocol is not a good solution
under a vulnerable environment. Although all of them did not intentionally provide any solution
for security issue, replication-based protocol performs slightly better in a highly stressed

68
environment. The overheads of prediction-based protocol are also acceptable, which indicates it is
able to handle the restricted hardware situation in DTNs.
6. CONCLUSION
Due to the property of Delay Tolerant Networks for no guaranteed continuous connectivity,
choosing a proper routing protocol and finding its optimal configuration are challenging
problems. In this paper, we evaluate some existing representative routing protocols for DTNs
through extensive trace-based simulations. Flooding-based routing protocols, such as Epidemic
and Spray and Wait, can achieve better delivery rate by sacrificing the memory, while prediction-
based routings, like PRoPHET and 3R, are more efficient when considering the delivery
overheads. Although all of the protocols we evaluate have rare capability to handle the hostile
attack from adversary users, the prediction-based is able to achieve a relative higher delivery rate.
This paper provides important guidance for routing protocols design. Our future work is to design
an efficient routing protocol for DTNs with high delivery rate and low delivery overheads.
REFERENCES
[1] T. Cao, X. Zhang, L. Kong, X. Liu, W. Shu, and M. Wu, “Traffic aware routing in urban vehicular
networks,” in WCNC, pp. 2004–2009, 2013.
[2] P. Juang, H. Oki, Y. Wang, M. Martonosi, L. S. Peh, and D. Rubenstein, “Energy-efficient computing
for wildlife tracking: Design tradeoffs and early experiences with zebranet,” SIGARCH Comput.
Archit. News, vol. 30, no. 5, pp. 96–107, 2002.
[3] Y. Zhang and J. Zhao, “Social network analysis on data diffusion in delay tolerant networks,” in
MobiHoc, pp. 345–346, 2009.
[4] A. Vahdat and D. Becker, “Epidemic routing for partially-connected ad hoc networks,” tech. rep.,
2000.
[5] T. Spyropoulos, K. Psounis, and C. S. Raghavendra, “Spray and wait: An efficient routing scheme for
intermittently connected mobile networks,” in WDTN, pp. 252–259, 2005.
[6] A. Lindgren, A. Doria, and O. Schel´en, “Probabilistic routing in intermittently connected networks,”
SIGMOBILE Mob. Comput. Commun. Rev., vol. 7, pp. 19–20, July 2003.
[7] J. Burgess, B. Gallagher, D. Jensen, and B. Levine, “Maxprop: Routing for vehicle-based disruption-
tolerant networks,” in INFOCOM, pp. 1–11, April 2006.
[8] Q. Yuan, I. Cardei, and J. Wu, “An efficient prediction-based routing in disruption-tolerant
networks,” IEEE Trans. Parallel Distrib. Syst, vol. 23, no. 1, pp. 19–31, 2012.
[9] L. Vu, Q. Do, and K. Nahrstedt, “3r: Fine-grained encounter-based routing in delay tolerant
networks,” in WoWMoM, pp. 1–6, 2011.
[10] A. Lindgren, A. Doria, E. Davies, and S. Grasic, “Probabilistic routing protocol for intermittently
connected networks,” Internet-Draft, 2012.
[11] J. Scott, R. Gass, J. Crowcroft, P. Hui, C. Diot, and A. Chaintreau, “CRAWDAD data set
cambridge/haggle (v. 2006-01-31).” http://crawdad.org/cambridge/haggle/, 2006.
[12] N. Eagle and A. (Sandy) Pentland, “Reality mining: sensing complex social systems,” Pers. Ubiq.
Comput., vol. 10, no. 4, pp. 255–268, 2006.
[13] D. Kotz, T. Henderson, and I. Abyzov, “CRAWDAD data set dartmouth/campus (v. 2004-12-18).”
http://crawdad.org/dartmouth/campus, 2004.
[14] J. Shen, S. Moh, and I. Chung, "Routing protocols in delay tolerant networks: A comparative survey,"
in Proceeding of 23rd International Technical Conference on Circuits/systems, Computer and
communications, pp. 1577-1580, 2008

69
Authors
Luming Wan, is a Master candidate from Department of Computer Science, University of Otago, New
Zealand. He obtained B.S. degree from University of Otago, New Zealand in 2012. His research interests
include Delay Tolerant Network (DTN), Social Network, etc.
Feiyang Liu, is a PhD candidate from Department of Computer Science, University of Otago, New
Zealand. He obtained B.S. and M.S. degrees of telecommunication engineering from Xidian University,
China in 2009 and 2012, respectively. His research interests include Network on Chip (NoC), Optical
Network on Chip (ONoC), Delay Tolerant Networks (DTN), etc.
Yawen Chen, Yawen Chen obtained her PhD degree in Computer Science from The University of
Adelaide in Australia in 2008. After her PhD study, she worked as postdoctoral researcher at Royal
Institute of Technology (KTH) in 2009. She has been a Lecturer in the University of Otago in New Zealand
since 2011. Her research interests include resource optimization and performance evaluation in computer
networking and computer architecture (optical networks, interconnection networks, green computing, and
etc).
Haibo Zhang, received the MSc degree in Computer Science from Shandong Normal University, China in
2005, and the PhD degree in Computer Science from the University of Adelaide, Australia in 2009. From
2009 to 2010, he was a postdoctoral research associate at Automatic Control Laboratory, KTH, Sweden.
Currently he is a lecturer at Computer Science department of University of Otago, New Zealand. His
research interests include real-time industrial wireless communications, wireless sensor/ad hoc networks,
delay-tolerant networks, green computing, distributed algorithms and protocol design.

ROUTING PROTOCOLS FOR DELAY TOLERANT NETWORKS: SURVEY AND PERFORMANCE EVALUATION

More Related Content

What's hot (20)

Viewers also liked (19)

Similar to ROUTING PROTOCOLS FOR DELAY TOLERANT NETWORKS: SURVEY AND PERFORMANCE EVALUATION (20)

Recently uploaded (20)

ROUTING PROTOCOLS FOR DELAY TOLERANT NETWORKS: SURVEY AND PERFORMANCE EVALUATION