Content uploaded by Song Guo
Author content
All content in this area was uploaded by Song Guo on Mar 04, 2016
Content may be subject to copyright.
A Multi-Attribute Decision Making Approach to
Congestion Control in Delay Tolerant Networks
Kaimin Wei∗, Song Guo†, Deze Zeng†and Ke Xu∗
∗State Key Lab of Software Development Environment, Beihang University, China
†School of Computer Science and Engineering, The University of Aizu, Japan
Abstract—DTNs are prone to congestion due to limited re-
source on each node. We aim to develop an effective congestion
control mechanism in this paper. For this purpose, we first
identify a list of major congestion factors by analyzing the causes
of congestion. We then model the congestion control as a multiple
attribute decision making problem (MADM), in which the weight
of congestion factors is measured by an entropy method. To solve
this problem, we develop a MADM-based congestion control
mechanism that determines the forwarding order of a set of
messages on each encounter event. Moreover, we design a buffer
management scheme that deletes messages whose removal would
incur least impact to the network performance when buffer
overflows. Extensive real-trace driven simulation is conducted
and the experimental results finally validate the efficiency of our
proposed congestion control mechanism.
Index Terms—Delay Tolerant Networks, congestion control,
multiple attribute decision making.
I. INTRODUCTION
Recently the pervasiveness of mobile communication de-
vices (e.g., smart phone) has increased significantly and the
possibility of a novel communication paradigm without an
existing fixed infrastructure has become a reality. Haggle [1]
and Sassy [2] show that data transmission among people in a
small-scale communication scenario (e.g., university campus)
can be successfully achieved by opportunistic encounters,
despite the absence of contemporaneous end-to-end paths. This
communications paradigm is often referred to as a Delay Tol-
erant Networks (DTNs) [3] [4], which is designed to operate
in the challenged networking environment characterized by
dynamic network topology and intermittent connectivity.
To overcome these communication constraints and sup-
port delay-tolerant applications, numerous routing protocols
[5] that forward messages using a “store-carry-and-forward”
scheme for DTNs have emerged. They typically assume that
nodes have unlimited resources (e.g., storage capacity), and
take part in the routing process of others, e.g., storing messages
in a local buffer for others. However, DTN nodes are of
limited storage capacity in reality. When intermediate nodes
cache messages for others, their storage space can be quickly
overwhelmed. As a result, it leads to congestion and ultimately
a reduced delivery ratio. Furthermore, given the lack of stable
end-to-end paths in DTNs, the message replication approach
is commonly used to increase the delivery ratio. The most
typical example is Epidemic [6], where multiple copies easily
deplete the limited storage resource available to DTN nodes.
This approach could cause severe congestion and message
loss, resulting in a low network efficiency. Thus, it is important
and necessary to develop a congestion control mechanism for
DTNs.
Slowing the message sending rate at sources and flow-
based feedback have been successful at preventing Internet
congestion [7]. However, the lack of stable end-to-end paths
and the whole network information in DTNs eliminate the
use of these approaches. Instead, congestion control in DTNs
becomes a problem of message dropping, forwarding, or repli-
cation, based on local and partial information. The majority of
previous research work in congestion control focus on message
dropping [8], which aims to discard those messages with
less help on network efficiency. Later, the recently emerging
work on congestion control has been concerned with adaptive
replication management [9] [10] and traffic distribution [11]
[12]. All these approaches are based on the global network
information, which is challenging for DTNs. A more attainable
and efficient congestion control approach is to respond quickly
to congestion, using only local information. Each node acts
autonomously, i.e., choosing forwarding messages which cause
a low probability of potential congestion, and drops messages
which have the least effect on the delivery ratio. The key
but tough part of this algorithm is to determine the most
appropriate messages that should be forwarded or dropped
with local information.
The goal of our research is to gain an understanding of
congestion behavior in social-like DTNs, where nodes exhibit
long-term regularity mobility law, and then to develop an
efficient local congestion control mechanism, which ultimately
leads to an improved network efficiency, e.g., an increased
delivery ratio, a reduced delivery overhead, or a shortened av-
erage delivery delay. More specifically, our congestion control
mechanism is implemented through quickly forwarding the
messages with small size, long TTL (time-to-live), and short
dwelling time at their receiving nodes with sufficient buffer
capacity. Given these targets, the main contributions of our
work are summarized as follows.
•We specify the main factors to congestion in DTNs and
discuss how to exploit them in the design of congestion
control mechanism.
•We model the congestion control problem as a multiple
attributes decision making problem (MADM). Based on
this, we develop a MADM-based congestion control
mechanism. To our best knowledge, this is the first work
that utilizes decision theory to cope with the congestion
problem in DTNs.
•To validate our proposal, we compare a number of rep-
resentative routing algorithms against their congestion-
aware version, incorporating our MADM-based con-
gestion control mechanism, by using real-trace driven
simulations. The evaluation results show that DADM-
based approach can effectively cope with congestion and
improve network efficiency.
The remainder of our paper is organized as follows. Section
II reviews the related work. Section III gives the network
model and the congestion model. Section IV mentions the
overview of our solution. Section V details the design of the
MADM-based congestion control mechanism. Section VI eval-
uates its performance by comparing the representative routing
algorithms with and without our MADM-based congestion
control mechanism. Finally, Section VII concludes our work.
II. RE LATE D WOR K
A. Congestion Control in DTNs
Congestion in DTNs can be classified into link congestion
and storage congestion. A link congestion occurs when two or
more nodes within transmission range of each other compete
with the same channel. A storage congestion happens only
when messages contend for the use of limited buffer. In the
remainder of this paper, we will use the term “congestion” to
refer to the “storage or buffer congestion” that most frequently
occurs in DTNs.
In DTNs, the majority of previous work in congestion
control is concerned with buffer management [8], which aims
to discard some appropriate messages when the buffer fills up.
Although these algorithms can alleviate the congestion level
of the network, the obtained gain comes at the expense of net-
work performance. Thus they are not the ideal solution to cope
with congestion. Later on, congestion control mechanisms
using replication management [9] [10] or traffic distribution
[12] [11] have attracted attention recently. Retiring Replicas
[10] is an adaptive replication protocol that adjusts a node’s
copy number based on the information of the whole network,
which is difficult to be obtained in DTNs. By estimating the
load and expected delays of a set of neighbors, CAFE [12]
chooses targets for offloading the messages from a potential
intermediate node that is about to get congested. Fair route [11]
aims to change the direction of data-flow to avoid congesting
popular nodes.
B. Routing in DTNs
Congestion control mechanism always works with routing
protocols, which have been studied extensively in DTNs. Most
of them utilize the message replication approach to overcome
communication challenges in DTNs. A famous example is
Epidemic [6], which floods many copies to all its neighbors
within transmission range so that the copies are quickly dis-
tributed through the network. Although Epidemic can achieve
the highest delivery rate and shortest average delay, it suffers
intolerable load due to its flooding nature. Later, numerous
routing algorithms have been proposed, e.g., Spray&Wait [13],
TABLE I
LIS T OF NOTATIO NS
Variable Description
ninode i
mjmessage j
cfkcongestion factor k
wkweight of congestion factor ak
vj,k the value of message mjfor congestion factor cfk
Smjsize for message mj
F Bnifree buffer size of node ni
Dni,mjnode delay for message mjin node ni
Umjutility vaule of message mj
Mniset of messages stored in ni
Dniset of node delay of messages in ni
Setni,njthe forwarding set of nito nj
Prophet [14], and others, to reduce the overhead of Epidemic.
Spray&Wait spreads multiple copies to the first encountered
nodes, relying on them to finish message transmission. Prophet
first estimates a probabilistic metric called delivery probability
to message’s detonation, and then forwards messages to the
nodes which have an increased delivery probability.
III. MOD EL
In this section, the network model and congestion model
are presented. Some important notations used in the rest of
the paper are summarized in Table I.
A. Network Model
We consider a social-like delay-tolerant network, in which
one node can directly communicate with another if they move
into the reciprocal radio communication range of each other.
A node can simultaneously serve as a source/destination of a
message and a relay for others’ messages. A message traverses
the network by being relayed from one node to another until it
reaches its destination. In such network, long-term regularity
mobility is usually exhibited. For example, some pairs of nodes
(e.g., acquaintances) consistently meet more frequently than
other pairs (e.g., strangers) over time. Hence the historical
contact information implies the future encounter opportunity.
This has been extensively verified [15].
All mobile nodes are resource constrained, e.g., with limited
storage space. When plenty of messages are generated and
traversed in the network, nodes are prone to congestion since
they have to store messages for others. Although messages
in DTNs are delay tolerant, they are usually with a specified
lifetime, which is often evaluated as the time-to-live (TTL)
since its generation. A message will be discarded when its
TTL expires.
B. Congestion Model
We consider the following congestion factors.
Free buffer size. A buffer with larger capacity implies that
it can store more messages, and thus the probability of the
occurrence of congestion is reduced. An extreme example is
that no congestion will happen in the node with infinite storage
space.
Message size. The larger message size, the smaller number
of stored messages. Consequently, the probability of conges-
tion is increased. Thus, it is reasonable to consider message
size when designing a efficient congestion control mechanism.
In our work, we prefer to forward the messages with small
size.
Message TTL. Messages in DTNs must be stored by
intermediate nodes for a period of time in the process of
routing. The shorter the stored message’s TTL, the higher the
probability that the message is discarded. This will result in a
waste of storage resource and transmission opportunity. Thus,
it is reasonable to consider message’s TTL when designing
an efficient congestion control mechanism. In our work, we
would also prefer to forward the messages with long TTL.
Message’s node delay. It refers to the time that a message
stays in a node. This factor determines the utilization of
storage space because the buffer is a reusable resource. The
shorter the node delay of stored messages, the more the
number of stored messages in a period of time, and thus the
higher the buffer utilization. It is crucial to take message’s
node delay into account when designing a congestion control
mechanism. However, it is difficult to get the node delay
of messages in a particular node because of lacking the
global network information and future encounter knowledge.
Fortunately, nodes in our network model exhibit the long-term
regularity mobility law. We therefore utilize historical average
inter-contact time to denote node delay, Dni,mjof message mj
in node nito the destination nd, which is defined as follows:
Dni,mj=Pl
k=1 fni,nd(k)
l,
where lis the total contact number between niand njin
the esplased time, and fni,nd(k)denotes the k-th inter-contact
time between nodes niand nd.
IV. OVERV IE W OF OU R PAOPOSAL
A. Probelm Statement
To overcome the communication challenges in DTNs, most
DTN routing algorithms work in the “store-carry-and-forward”
way, in which intermediate nodes have to store others’ mes-
sages in its limited buffer for a period of time. However,
storage resource is constrained for nodes. Consequently a mass
of messages traversing the network will aggravate node load
and further give rise to congestion. Moreover, most routing
algorithms utilize a message replication approach to generate
multiple copies of the same message and inject them into the
network to improve the delivery ratio. These copies can easily
overwhelm node buffer, which is prone to lead to congestion
and eventually reduce network efficiency.
In this paper, we aim to design an efficient congestion
control mechanism, which could cope with the congestion
problem without causing any side effect to network. In partic-
ular, we study the following two subproblems to achieve our
target.
Selective Forwarding Set. It is to determine a set of
candidate forwarding messages and their forwarding order. In
congestion control, the ideal forwarding set should include
the messages, that are most likely to alleviate congestion to
the recipient, with higher forwarding priority. Nevertheless,
the congestion process in DTNs is very complicated, and
congestion is always caused by several factors comprehensive-
ly. Moreover, the global network information is not always
availably. Thus, it is challenging to identify these messages
with partial information.
Buffer Management. It determines which messages should
be discarded when the buffer overflows. Efficient buffer man-
agement is to delete the messages with the least effect on
the overall network performance. However, it is difficult to
find these messages because of lacking of the global network
information.
B. Methodology
The congestion process is very complicated, and is always
induced by several factors but not a simple one. An effective
way to congestion is to figure out the role of different
congestion factors in the process of congestion, and then give
them corresponding weights. The key is to measure the weight
of congestion factors.
Numerous previous work in other fields has proved that
the multiple attribute decision making method [16] is an
ideal and mature way in decision-making environments. It is
concerned with structuring and solving decision and planning
problems involving multiple attributes. Since each attribute has
a different meaning, it cannot be assumed that they all have
equal weights, and as a result, finding the appropriate weight
for each attribute is a key part. Various methods for measuring
weights can be categorized into two groups: subjective and
objective. Compared with subjective methods, the weights
measured are more reliable since they are not affected by
the preference of decision makers. The most typical objective
approach is entropy proposed by Shannon [17].
Inspired by this, we exploit the properties of the multiple
attribute decision making way to address control congestion
problem in DTNs through identifying the messages which have
the least effect on congestion.
C. Solution Overview
Here we mainly give an overview of our MADM-based
congestion control mechanism and explain how it works. The
overall architecture of MADM-based is shown in Fig.1, in
which routing modular and congestion control module work
together to make forwarding decisions for messages in the
buffer. During node contact, each module exchange status
data with its peer: the routing modules exchange routing
information such as delivery probability, while the congestion
control modules exchange node buffer statistics, e.g., free
buffer space, the node delay of messages. In addition, each
node acts indecently relying on local information only, to avoid
to forward messages to the congested nodes.
After the routing module chooses the encountered node as
a relay, congestion control mechanism is triggered. Different
from routing that aims to choose the most appropriate nodes
Free$Buffer$Size
Rou-ng$Info
Forwarding$Msgs
Rou-ng$
Mechanism
Forwarding$
Set
Conges-on$
Control$
Mechanism
Buffer$
Management
Rou-ng$
Mechanism
Forwarding$
Set
Conges-on$
Control$
Mechanism
Buffer$
Management
Node Delay,
Fig. 1. The overall structure of our MADM-based congestion control
mechanism
as relay, the congestion control mechanism that we focus is
to determine the forwarding set and its transmitting order by
forwarding the messages that are likely to alleviate congestion
to recipient in the future.
Congestion control in DTNs is closely related to buffer
management. When the buffer overflows, it will delete the
messages with the least effect on the overall performance of
the network from the buffer.
V. MADM-BASED CONGESTION CONTROL
In this section, we model the congestion control problem
as a multiple attributes decision making problem to determine
the forwarding set and its transmission order. After that, we
propose a MADM-based congestion control mechanism.
A. Weight Determination
In the congestion factors considered, some are associated
with a particular message, e.g., message size, message TTL
and message’s node delay, while others are the same for all
messages, e.g., free buffer size. Therefore, we only consider
the former three factors, referred to as cf1,cf2, and cf3,
respectively, in the following when determining their weights.
We utilize an entropy [17] method to decide the weight of
each attribute. The detailed steps are outlined in the following.
Moreover, the congestion factor free buffer size will utilize to
determine the maximum number of messages in the forward-
ing set.
Step (1). Build an attribute matrix described as follows:
m1m2m3
cf1
cf2
cf3
v1,1
v1,2
v1,3
v2,1
v2,2
v2,3
v3,1
v3,2
v3,3
..
..
..
where rows and columns represent the congestion factor and
the message id, respectively. Each entry vi,j in the matrix is
the value of message miwith respect to factor cfj.
Step (2). Normalize the matrix by
pi,j =vi,j
Pn
j=1 vi,j
, i ∈[1,3], j ∈[1, n].
The raw data are normalized to eliminate anomalies in differ-
ent measurements. This process transforms different scales and
units among various criteria into common measurable units to
allow comparisons of different criteria.
Step (3). Compute entropy eias
ei=−e0
n
X
j=1
pi,j ln pi,j , i ∈[1,3],
where e0is the entropy constant and is equal to (ln m)−1.
Step (4). Measure the degree of diversification of entropy
as
di= 1 −ei, i ∈[1,3].
Step (5). Compute the weight wiof congestion factor cfi
as
wi=di
P3
k=1 ek
, i ∈[1,3].
B. Forwarding Set Determination
In order to measure the effect of congestion incurred by
a message to a specific node, we introduce a new metric
called message utility, to characterize such probability. The
larger the utility value, the smaller the congestion probability.
By considering various congestion factors in a comprehensive
way, the utility of message mi, denoted as Umi, is defined as
the sum of all congestion factors, i.e.,
Umi=
3
X
k=1
wk×vi,k,
where Umiis the utility value of message mi.
Instead of deciding how many messages should be included
into the forwarding set because of lacking the knowledge about
current contact duration, here we utilize a greedy approach to
select messages. The basic idea is to circularly put the message
with the largest utility value into the forwarding set until the
free buffer capacity is depleted at the recipient node.
Furthermore, due to the uncertainty of contact duration,
messages in the forwarding set may not be all sent out.
Therefore, it is important to determine the forwarding order.
We forward messages in a descending order of their values
until the contact disappears or no message is left in the
forward set. In this way, the total utility value of the messages
transferred is always maximized, and thus the probability of
congestion caused by transferring messages to the receiver is
expected to be minimized.
C. MADM-based Design
Based on the above work, we develop a MADM-based
congestion control mechanism, which models the congestion
control problem as a multiple attributes decision making prob-
lem. This mechanism works together with routing modules
to complete data transmission. After a node determines to
forward messages to the encountered node, DADM-based is
triggered. The detailed procedures at node niwhen nodes ni
and njmeet are summarized in Algorithm 1.
Algorithm 1 MADM-based Congestion Control at ni
Require: F Bnj,Mni,Dni,Dnj
Ensure: Setni,nj
1: exchange buffer state and node delay information
2: initialize the forwarding set Setni,nj
3: measure the weight of congestion factors by using the
approach introduceted in Section V-A
4: compute the utility vaule of messages in Setni,njby the
metric introducted in Section V-B
5: sort Setni,njin a descending order of message utility
6: for message mkin Setni,njdo
7: if Smk< F Bnjthen
8: F Bnj=F Bnj-Smk
9: else
10: Setni,nj=S etni,nj\ {mk}
11: end if
12: end for
The two nodes first exchange their buffer state and node
delay information. The forwarding set Setni,njis then initial-
ized by including any message that must satisfy the following
two conditions: 1) its node delay is shorter than in njthat in
ni, and 2) its node delay is shorter than its TTL. After that, ni
will measure the weight of congestion factors and compute the
utility value of all messages in Setni,nj, as revealed by lines
3-4 of Algorithm 1. Subsequently, niwill update Setni,nj
to ensure that the sum size of messages in Setni,njis not
greater than the free buffer size of node ni, denoted as F Bnj,
as shown in lines 5-12 of Algorithm 1. This could ensure that
the total utility value of the messages transferred is always
maximized. After Setni,njis finally determined, nistarts to
transmit messages until the connection is interrupted or no
message in Setni,njis left.
D. Buffer Management
Owing to the buffer capacity limitation, it is impractical
to keep all received messages in its local buffer. Therefore,
we shall consider a buffer management scheme to determine
which messages should be deleted when the buffer fills up.
The target is to drop those messages that have less impact
on the overall performance of the network, e.g., end-to-end
delay. More specifically, we adopt a message utility-based
buffer management, in which the message with the lowest
utility value is dropped firstly to ensure that the total utility
value of the messages buffered is always maximized. Note
that, unlike MADM-based congestion control, only the node
delay of the current node is considered to measure the weight
of congestion factors in our buffer management.
VI. PERFORMANCE EVALUATIO N
A. Simulation Setup
To evaluate the performance of MADM-based congestion
control mechanism, we implement it in a widely-used DT-
N simulator ONE [18]. Two practical mobility traces from
datasets Infocom2006 [1] and Sassy [2] are incorporated into
TABLE II
SIMULATIONS PARA ME TER S
Parameter Value
TTL (Time-to-live) 300 minutes
Warmup time 30% of the whole simulation time
Transmit speed 300 kBps
Range of Message size 20 80 kB
Node buffer size (default) 6 M
Message creation interval (default) 30 seconds
Number of copies in Spray&Wait 6
Spray way in Spray&Wait binary mode
Initial aging time unit in Prophet 30 seconds
ONE. We believe that these datasets cover a rich diversity of
network environments, from small conference (Infcom2006)
to spacious campus (Sassy), with a duration of several days
(Infocom2006) to more than two months (Sassy). Without loss
of generality, we consider the same buffer size at each node.
Messages are generated periodically at the source and their
destination is selected uniformly at random from the entire
whole. The message size is also uniformly distributed from
a range. Detailed simulation parameters in our simulated are
listed in Table II.
We consider three representative routing algorithms, Epi-
demic [6], Spray&Wait [13] and Prophet [14], in our experi-
ment. We also implement their congestion-aware version by in-
corporating our MADM-based congestion control mechanism,
called Epidemic-c, Spray&Wait-c and Prophet-c, respectively.
By comparing the original version of these algorithms against
their congestion-aware version, we can check the effectiveness
of our MADM-based mechanism. In addition, we use the
following performance metrics.
•Delivery Ratio: the ratio between the number of mes-
sages that are successfully delivered to their destination
nodes within their lifetime to the total number of mes-
sages generated.
•Delivery Overhead: the total number of forwardings
until all generated messages to be successfully received.
•Delivery Delay: the average time required to deliver a
message to its destination node since its generation time.
B. On the Effect of Node Buffer Size
In order to mimic different congestion levels of the network,
we change the size of the buffer, and then evaluate the perfor-
mance of our MADM-based congestion control mechanism in
various congestion environments.
1) Delivery Ratio Comparison: Fig. 2 shows the delivery
ratio of all algorithms on various data sets, as a function
of buffer size. The larger the buffer capacity, the higher the
delivery ratio. The reason behind is that the buffer size of
a node determines the maximum number of messages to be
buffered. Larger buffer space means more messages will be
stored in intermediate nodes, and thus these messages have
more chance to be delivered to their destination.
We can also see clearly from Fig. 2 that all congestion-
aware algorithms are of superior performance to the original
ones under different scales of buffer size. Taking data set
2M 4M 6M 8M 10M
0
0.2
0.4
0.6
0.8
1
Buffer Size
Delivery Ratio
Spray&Wait−c
Spray&Wait
Prophet−c
Prophet
Epidemic−c
Epidemic
(a) Infocom2006
2M 4M 6M 8M 10M
0
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
Buffer Size
Delivery Ratio
Spray&Wait−c
Spray&Wait
Prophet−c
Prophet
Epidemic−c
Epidemic
(b) Sassy
Fig. 2. Performance evaluation results of the delivery ratio
2M 4M 6M 8M 10M
0
0.2
0.4
0.6
0.8
1
Buffer Size
Delivery Overhead Percentage
Spray&Wait−c
Spray&Wait
Prophet−c
Prophet
Epidemic−c
Epidemic
(a) Infocom2006
2M 4M 6M 8M 10M
0
0.2
0.4
0.6
0.8
1
Buffer Size
Delivery Overhead ercentage
Spray&Wait−c
Spray&Wait
Prophet−c
Prophet
Epidemic−c
Epidemic
(b) Sassy
Fig. 3. Performance evaluation results of the delivery overhead
Infocom2006 as an example, we observe that Epidemic-c
outperforms Epidemic by 69% and 39% when the buffer
size is 2Mand 4M, respectively, as shown in Fig. 2(a).
We attribute the advantage of the congestion-aware algorithm
over the original one to three factors. Firstly, congestion-
aware algorithms only forward messages to those nodes that
could provide sufficient space to buffer incoming messages.
This could avoid message loss caused by no available storage
space at the receiver. However, all original algorithms fail
to achieve it. Secondly, when a contact opportunity occurs,
congestion-aware algorithms only forward those messages that
are most unlikely to incur congestion to the recipient node,
while the original ones do not consider it. Thirdly, congestion-
aware algorithms use a reasonable buffer management by
taking into account of network context, i.e., message dropping
decisions will be made when a node is about to get congested.
Nevertheless, original algorithms all overlook the importance
of buffer management and utilize a simple buffer management,
e.g., random drop.
In addition, another interesting phenomenon noticed from
Fig. 2 is that the superiority of congestion-aware algorithms
become smaller with the increasement of buffer size. This
indirectly proves that the effectiveness of our congestion
mechanism, because the buffer size of nodes implies the
congestion level of the network. When the buffer size of nodes
in networks increases, the congestion level of the network
reduces.
2) Delivery Overhead Comparison: Fig. 3 reveals the eval-
uation results on the delivery overhead of all routing algorithm
on different values of buffer size. Due to the large span of
the delivery overhead of different algorithms, it is difficult to
show their absolute values on one figure clearly. Instead, we
2M 4M 6M 8M 10M
3500
4000
4500
5000
5500
Buffer Size
Average Delivery Delay
Spray&Wait−c
Spray&Wait
Prophet−c
Prophet
Epidemic−c
Epidemic
(a) Infocom2006
2M 4M 6M 8M 10M
6000
6200
6400
6600
6800
7000
7200
Buffer Size
Average Delivery Delay
Spray&Wait−c
Spray&Wait
Prophet−c
Prophet
Epidemic−c
Epidemic
(b) Sassy
Fig. 4. Performance evaluation results of the average delivery delay
show the results in delivery overhead percentage, i.e., a ratio of
overhead generated by the original algorithm or its congestion-
aware version over their sum, to better illustrate the delivery
overhead of all algorithms. In other words, if the delivery
overhead percentage of a congestion-aware algorithm is less
than 50%, this congestion-aware algorithm has an improved
performance in terms of delivery overhead, comparing with
the original algorithm.
As shown in Fig. 3, it is obvious that all congestion-aware
algorithms perform better than their original ones in terms of
delivery overhead as the delivery overhead percentage of all
congestion-aware algorithms below the black dotted line. The
most typical example is Epidemic-c in Infocom2006. Delivery
overhead of Epidemic-c is 3.4% and 5.5% of Epidemic when
the buffer size is 2Mand 4M, respectively. Recall that we
achieve such high performance with the lowest overhead ratio
as shown in Fig. 2. This advantage results come from two
reasons: 1) MADM-based congestion control mechanism; and
2) message utility-based buffer management.
Furthermore, we can observe from Fig. 3 that the perfor-
mance gap between the congest-aware version and the original
one of Epidemic is much larger than that of Spray&Wait
and Prophet. This is caused by Epidemic’s flooding nature.
Epidemic spreads numerous copies upon every contact oppor-
tunity, which significantly aggravates node load, and in turn
leads to message loss. However, other two algorithms only
generate limited replicas into the network. The congestion
level of the network is much lower, and the improvement space
of the delivery overhead is narrowed.
3) Delivery Delay Comparison: Fig. 4 gives the evalu-
ation results on the delivery delay of all the protocols on
different data sets. It is clearly seen that all congestion-aware
algorithms have a shorter delivery delay than the original
ones. This phenomenon comes from three factors. The first
is that congestion-aware algorithms forward messages to the
node with a relatively short node delay. The second is that
congestion-aware algorithms only deliver messages to the node
that has sufficient storage space for newly arrived messages.
This could avoid message loss caused by overflow, which will
induce the source node to retransmit the original data. The
third is the buffer management ulitized in congestion-aware
algorithms that only discards the messages with the least effect
on network performance.
20s 25s 30s 35s 40s 45s
0
0.2
0.4
0.6
0.8
1
Message Generated Interval
Delivery Ratio
Spray&Wait−c
Spray&Wait
Prophet−c
Prophet
Epidemic−c
Epidemic
(a) Delivery Ratio
20s 25s 30s 35s 40s 45s
0
0.2
0.4
0.6
0.8
1
Message Generated Interval
Delivery Overhead Percentage
Spray&Wait−c
Spray&Wait
Prophet−c
Prophet
Epidemic−c
Epidemic
(b) Delivery overhead
20s 25s 30s 35s 40s 45s
4000
4500
5000
5500
6000
6500
Message Generated Interval
Average Delivery Delay
Spray&Wait−c
Spray&Wait
Prophet−c
Prophet
Epidemic−c
Epidemic
(c) Delivery Delay
Fig. 5. Performance evaluation results of the average delivery delay
C. On the Effect of Message Generation Interval
We also mimic various congestion levels of the network
by changing the message generation interval when the buffer
size is constant. The smaller interval, the more messages
generated, and thus the higher congestion level of the network.
We evaluate the performance of MADM-based mechanism
in different congestion environments, and similar simulation
results are obtained as shown in Fig. 5. Compared with the
original algorithm, the congestion-aware one has the improved
network performance in terms of delivery ratio, delivery over-
head and delivery delay. This is attributed to the advances
of our proposed congestion control mechanism, which can
effectively cope with the congestion problem without giving
any side effect to network.
In addition, compared with the original algorithms, the
benefits of congestion-aware algorithms is not evident when
the message generation interval becomes larger. The reason
behind is that a larger interval implies that fewer messages will
be created and injected into the network and the congestion
level of the network is also reduced. Thus, the benefits of the
congestion-aware algorithms are limited.
VII. CONCLUSION
In this paper, we aim to develop a congestion control
mechanism which can address the congestion problem without
bringing any side effect to network. We first analyze the cause
of congestion and find the congestion factors. Then, we model
the congestion control problem as a multiple attribute decision
making problem, so as to identify the messages with the least
potential congestion effect to the recipient. The weight of
congestion factors is measured by an entropy method, and the
congestion effect of each message stored to a specific node is
evaluated by the metric called message utility. After that, we
present a MADM-based congestion control mechanism, which
aims to decide the forwarding set. Moreover, we also present
a utility-based buffer management which considers the context
of nodes and messages. Extensive real-trace driven simulation
results finally validate the efficiency of our proposed conges-
tion control mechanism.
REFERENCES
[1] J. Scott, R. Gass, J. Crowcroft, P. Hui, C. Diot, and A. Chaintreau,
“Data Set Cambridge/haggle/imote/infocom2006,” Downloaded from
http://crawdad.cs.dartmouth.edu/cambridge/haggle/imote/infocom2006,
May 2009.
[2] G. Bigwood, D. Rehunathan, M. Bateman, T. Hen-
derson, and S. Bhatti, “data set st andrews/sassy,”
http://crawdad.cs.dartmouth.edu/standrews/sassy, jun 2011.
[3] Y. Cao and Z. Sun, “Routing in Delay/Disruption Tolerant Networks:
A Taxonomy, Survey and Challenges,” IEEE Communications Surveys
and Tutorials, vol. 15, no. 2, pp. 654–677, 2013.
[4] H. Zhu, M. Dong, S. Chang, Y. Zhu, M. Li, and X. Shen, “ZOOM:
Scaling the mobility for fast opportunistic forwarding in vehicular
networks,” in INFOCOM, 2013, pp. 2832–2840.
[5] K. Wei, X. Liang, and K. Xu, “A Survey of Social-Aware Routing
Protocols in Delay Tolerant Networks: Applications, Taxonomy and
Design-Related Issues,” IEEE Communications Surveys and Tutorials,
no. 99, 2013.
[6] A. Vahdat and D. Becker, “Epidemic Routing for Partially-connected
Ad Hoc Networks,” Duke University, Tech. Rep., 2002.
[7] S. Floyd and V. Jacobson, “Random early detection gateways for
congestion avoidance,” Networking, IEEE/ACM Transactions on, vol. 1,
no. 4, pp. 397–413, 1993.
[8] A. Krifa, C. Barakat, and T. Spyropoulos, “Message Drop and Schedul-
ing in DTNs: Theory and Practice,” Mobile Computing, IEEE Transac-
tions on, vol. 11, no. 9, pp. 1470–1483, 2012.
[9] A. Balasubramanian, B. Levine, and A. Venkataramani, “Replication
Routing in DTNs: A Resource Allocation Approach,” Networking,
IEEE/ACM Transactions on, vol. 18, no. 2, pp. 596–609, 2010.
[10] N. Thompson, S. Nelson, M. Bakht, T. Abdelzaher, and R. Kravets,
“Retiring Replicants: Congestion Control for Intermittently-Connected
Networks,” in INFOCOM, 2010, pp. 1–9.
[11] J. Pujol, A. Toledo, and P. Rodriguez, “Fair Routing in Delay Tolerant
Networks,” in INFOCOM, 2009, 2009, pp. 837–845.
[12] M. Radenkovic and A. Grundy, “Congestion aware data dissemination
in social opportunistic networks,” SIGMOBILE Mob. Comput. Commun.
Rev., 2010, vol. 14, no. 3, pp. 31–33, Dec. 2010.
[13] T. Spyropoulos, K. Psounis, and C. Raghavendra, “Efficient Routing in
Intermittently Connected Mobile Networks: The Multiple-Copy Case,”
Networking, IEEE/ACM Transactions on, vol. 16, no. 1, pp. 77–90, 2008.
[14] A. Lindgren, A. Doria, and O. Schel´
en, “Probabilistic routing in inter-
mittently connected networks,” SIGMOBILE Mob. Comput. Commun.
Rev., vol. 7, no. 3, pp. 19–20, Jul. 2003.
[15] Q. Yuan, I. Cardei, and J. Wu, “An Efficient Prediction-Based Routing in
Disruption-Tolerant Networks,” Parallel and Distributed Systems, IEEE
Transactions on, vol. 23, no. 1, pp. 19–31, 2012.
[16] J.-B. Yang and G. Singh Madan, “An evidential reasoning approach for
multiple-attribute decision making with uncertainty,” Systems, Man and
Cybernetics, IEEE Transactions on, vol. 24, no. 1, pp. 1–18, 1994.
[17] C. E. Shannon, “A mathematical theory of communication,” SIGMO-
BILE Mob. Comput. Commun. Rev., vol. 5, no. 1, pp. 3–55, Jan. 2001.
[18] A. Ker¨
anen, J. Ott, and T. K¨
arkk¨
ainen, “The ONE Simulator for DTN
Protocol Evaluation,” in Proceedings of the 3rd International ICST
Conference on Simulation Tools and Techniques (SIMUTools), 2009.