Available via license: CC BY 4.0
Content may be subject to copyright.
International Journal of Computer Networks & Communications (IJCNC) Vol.6, No.3, May 2014
DOI : 10.5121/ijcnc.2014.6302 13
LINK-AWARE NICE APPLICATION LEVEL
MULTICAST PROTOCOL
Dina Helal
1
, Amr Naser
1
, Mohamed Rehan
1
, and
Ayman El Naggar
2
1
Intel Labs, Cairo, Egypt
2
Faculty of Computer Science and Information Technology, German University in Cairo,
Cairo, Egypt
A
BSTRACT
Multicast is one of the most efficient ways to distribute data to multiple users. There are different types of
Multicast such as IP Multicast, Overlay Multicast, and Application Layer Multicast (ALM). In this paper,
we present a link-aware Application Layer (ALM) Multicast algorithm. Our proposed algorithm, Link
Aware-NICE (LA-NICE) [1], is an enhanced version of the NICE protocol [2]. LA-NICE protocol uses the
variations of bandwidth or capacity in communication links to improve multicast message delivery and
minimize end-to-end delay. OMNeT++ simulation framework [3] was used to evaluate LA-NICE. The
evaluation is done through a comparison between LA-NICE and NICE. The simulation results showed that
LA-NICE produces an increased percentage of successful message delivery ranging from 2% to 10%
compared to NICE. Also, LA-NICE has less average delay and less average message hop count than NICE
which reduces the overall latency of message delivery.
K
EYWORDS
Application Level Multicast, Multicast tree, Overlay networks, Link Aware.
1.
I
NTRODUCTION
As the number of Internet users increase, data delivery over the Internet becomes more
challenging as networks get more overloaded and congested. Currently, data exchange through
the Internet is mainly based on unicast (point-to-point between two computers). So, if millions of
users try to stream an important broadcast event like the world soccer cup, instead of broadcasting
the data to all users, the data source sends a copy of the data to each of the users so the source
keeps transmitting the same packet a million times. This leads to redundant traffic in the network
in addition to overloading the data source resulting in inefficient data delivery and an increase in
packet loss. Multicast was introduced as an alternative to unicast in such cases. In multicast, the
source send contents to a sub-server set and each one of those sub-server set forward the content
to a different group of users. There are several types of multicast such as IP, overlay, and
application level. In IP Multicast, the multicast process is implemented at the IP level during
packet transmission. IP multicast provides an efficient multicast technique. However, it was never
widely deployed in the Internet due to multiple reasons including the fact that it requires changes
at the infrastructural level which slows down the pace of deployment. Also IP multicasting
introduces high complexity and serious scaling constraints at the IP layer in order to maintain a
state for each multicast group. As a result of the non-acceptance of IP Multicast, the Application
layer multicast (ALM) approach was proposed. ALM, also called End-System Multicast, was
proposed as an alternative implementation of the multicast technique to the IP Multicast
implementation. ALM builds a virtual topology on top of the physical Internet to form an overlay
network. Each link in the virtual topology is a unicast link in the physical network [4]. Therefore,
International Journal of Computer Networks & Communications (IJCNC) Vol.6, No.3, May 2014
14
the IP layer provides a unicast datagram service, while the overlay network implements all the
multicast functionality such as dynamic membership maintenance, packet duplication and
multicast routing [5].
The rest of the paper is organized as follows. Section 2 describes NICE ALM protocol. Section 3
describes the enhanced Link-Aware NICE protocol. Section 4 presents the simulation design, the
implementation and the evaluation and test results. Finally, section 5 presents the conclusion
based on the simulation results.
2.
ALM OVERLAY CONSTRUCTION
There have been a number of approaches to design ALM protocols. In this section, some of these
approaches are introduced. To design a protocol some factors need to be taken into consideration
regarding how to manage nodes in multicast groups. This includes how users find out about
multicast groups, how they join a group, how they leave the group, abruptly or after notifying the
rest of the group, how to handle partitions and splits in the group. These factors of the group
management mechanism depend on the nature of the application.
In theory, the overlay network can be viewed as a fully connected graph, as each node can reach
every other node in the network via unicast connections. However, only a small subset of the
overlay links should be included in the ALM overlay. Hence, the basic functionality of an ALM
overlay construction technique is to identify these links and maintain the connectivity of the
overlay. The resultant overlay can be in the form of a tree or a mesh which serves two purposes:
• Control topology: is a mesh which provides redundant paths between the members
• Data distribution topology: Which is usually a tree used to ensure loop-free routing
A typical overlay construction technique consists of two phases, the joining phase and the
maintenance phase. The joining phase refers to the process where a newcomer is joining an
overlay. The maintenance phase manages the connectivity of the overlay. The ALM overlay can
be built in a centralized or a decentralized manner.
2.1. Centralized Approach
In the centralized approach, the overlay creation and management are performed by a central
controller. Newcomers can learn about the identity of the controller from the Rendezvous Point
(RP) that acts as a query server to provide existing members’ information to newcomers. The
controller uses full knowledge of all members, and potentially the performance metrics (e.g. delay
and bandwidth) between the members, to compute a high quality overlay. The computed overlay
structure is then distributed to the members in the form of the neighboring relationships (i.e.
links) between the nodes.
On receiving such information, the members initiate connections to their assigned neighbors.
Once attached to the overlay, the members enter the maintenance phase where they monitor the
connections with their respective neighbors. Any changes to the neighbors’ status (e.g. leave/fail)
gets reported such that the overlay can be repaired.
2.2. Decentralized Approach
In the decentralized approach, the overlay creation and management are done by the members in a
distributed fashion. Consider newcomer node X, first it requests a list of members in the multicast
International Journal of Computer Networks & Communications (IJCNC) Vol.6, No.3, May 2014
15
group from the RP. From the given list, X then selects one or more members as joining targets and
sends to each of them a joining request message. When a node, receives a request message, it
performs an admission control decision for the requesting node. The main decision criterion is
whether it has spare capacity for a new link. Once joined on the overlay, X enters the maintenance
phase and begins to monitor the status of its neighbors. Any changes to the neighbors’ status is
normally handled by X itself (i.e. the affected member).
While the centralized approach simplifies overlay construction and management, it may not scale
well. In particular, the central controller needs to keep track of the information about all members,
which is highly dynamic. In addition, it creates a single point of failure problem. Hence,
decentralized solutions are more preferable. Decentralized protocols are classified into two
groups: Mesh-first and Tree-first.
1) Mesh-First: In the Mesh-first approach, members keep a connected mesh topology among
themselves as shown in Figure 1. Usually the source is chosen as a root and a routing algorithm is
run over the mesh relative to the root to build the tree [6]. This mesh topology is explicitly created
at the beginning, hence it is known. On the other hand, the resulting tree topology is unknown. So
the quality of the tree depends on the quality of the mesh chosen. The advantage of Mesh-first
approach is that it gives more freedom to refine the tree. It is possible to manipulate the tree
topology to a significant extent by selecting mesh neighbors and changing the metrics. A Mesh-
first approach is therefore more robust and responsive to tree partitions and is more suitable for
multi-source applications, at the cost of higher control overhead. Mesh-first can either be
unstructured or structured.
Figure 1. Mesh-first design
Unstructured Mesh: An unstructured mesh is a graph where no special structural information can
be inferred from the graph to aid routing and management. Forming such a topology is simple: a
newcomer simply attaches itself to some randomly selected members. To obtain loop-free routing
trees from this type of topology, a conventional routing protocol such as distance-vector or link-
state is needed.
Structured Mesh: While routing in an unstructured mesh is difficult, it is possible to relate the
mesh members in a structural manner to simplify routing and management. One of the commonly
used structural meshes is distributed hash table (DHT).
International Journal of Computer Networks & Communications (IJCNC) Vol.6, No.3, May 2014
16
2) Tree-First: In the Tree-first approach, the tree is built directly without any mesh. The
members explicitly select their parent from the known members in the tree. This may require
running an algorithm to detect and avoid loops. The reason for using the Tree-first approach over
the Mesh-First approach is that the Tree-First approach gives direct control over the tree. This
control is valuable for different aspects such as selecting a best parent neighbor that has enough
resources, or responding to the failed members with a minimum impact to the tree. Another
advantage of the Tree-First approach is independent actions from each member. It makes the
protocol simple as it has a lower communication overhead. For example, when a member changes
a parent, it drags all of its descendants with it. This is desirable in the sense that the descendants
do not need to change their neighbors; in fact, they are unaware of the incident. However, this can
also result in uneven and less efficient trees.
2.3. Existing ALM Protocols
Some of the most popular existing ALM protocols introduced had different designs. As shown in
Figure 2, CoopNet [7] an example of a Centralized ALM protocol. NICE [2] and Zigzag [8] as
examples for centralized Tree-first protocols. Narada [11] is an example of unstructured mesh and
finally Scribe [9] is an example of structured mesh using DHT.
1) CoopNet [7] is a centralized ALM protocol that provides a resilient technique to deliver
streaming contents from a single source. It constructs multiple distribution trees, across the range
of given nodes, transmitting different multiple description coding MDC) descriptions on every
tree. MDC is a method of encoding audio and/or video signal into separate streams, or
descriptions, such that any subset of these descriptions can be received and decoded into a signal
with distortion (with respect to the original signal) depending on the number of descriptions
received, the more descriptions received, the better the quality of the reconstructed signal.
In CoopNet, each sub-stream is delivered using a different distribution tree (formed by the same
set of members). This delivery mechanism provides robustness as the probability of all streams
concurrently to fail to arrive is very low.
2) Zigzag [8] adopts a similar hierarchical structure to NICE for overlay maintenance. In
NICE, the cluster-leaders manage data forwarding to their children which leads to heavier load on
higher level nodes. ZIGZAG reduces the bottleneck at the higher level nodes by increasing the
size of the cluster in the top layer and sharing the cluster leader load with all the nodes in the
cluster [10].
International Journal of Computer Networks & Communications (IJCNC) Vol.6, No.3, May 2014
17
Figure 2. ALM protocols
3) Narada [11] is a protocol suitable for small and sparse groups, for audio/video conferencing,
virtual teaching and multiplayer games, not for the delivery applications. The routing control
plane of Narada is an unstructured mesh. According to different sending sources, the
corresponding data delivery trees are constructed.
In order to obtain a high-quality tree, constructing a good mesh is very important. If a new
member M wishes to join the group G, first of all, it will get a list of group members. The list
must contain at least one currently active group member. Then, M randomly selects a few group
members from the list available to it and sends them messages requesting to be added as a
neighbor. This process is repeated until it gets a response. Subsequently, M starts exchanging
refresh messages with its neighbors. At this time, it has successfully joined the group.
Narada runs a distance vector protocol on top of the mesh. The source trees are constructed by the
reverse shortest path between each recipient and the source. Different sources lead to different
data delivery tree. A member M that receives a packet from source S through a neighbor N
forwards the packet only if N is the next hop on the shortest path from M to S.
One important issue in mesh management is the partition problem. Detecting a partition in a mesh
is harder than in a tree. To solve the problem, Narada requires each node to maintain the
membership of the overlay, where the dissemination of the membership is integrated with the
routing protocol. This, however, leads to a relatively high control overhead. Thus, Narada is
effective only for small-scale applications.
4) Pastry is an overlay and routing network for the implementation of a distributed hash table
(DHT). In Pastry, each overlay node is assigned a random node ID that is uniformly selected from
a one-dimensional circular namespace of 128 bits. Given a message and a destination node ID,
Pastry routes the message to a node with the node ID that is numerically closest to the key, among
all live nodes. Figure 3 illustrates an example where a message targeted for node c is routed from
node a to node b. The Pastry overlay is created in the following order.
International Journal of Computer Networks & Communications (IJCNC) Vol.6, No.3, May 2014
18
• First, a new member uses a deterministic hash function to create a unique node ID.
• It then sends a join message addressed to its own node ID through a close-by overlay node,
say node a. Pastry assumes that the knowledge of such a node is learnt from some out-of-band
techniques, such as an expanded ring search.
• Node d then forwards the message hop-by-hop towards the destination node ID.
• Finally, the message reaches node b which has the closest node ID to e. The newcomer
obtains the state information from nodes on the path from a to f to establish its routing table
and neighbors list.
5) Scribe [9, 12] is a network of Pastry nodes, where each node runs the Scribe application
software. Any Scribe node can create a topic and other nodes can register their interest in this
topic by becoming members of the topic. A Scribe node with the appropriate credentials for the
topic can then publish events related to the topic and Scribe will disseminates these events to the
topic’s subscribers. Scribe uses best-effort dissemination of events and does not guarantee
ordered delivery. However reliability can be implemented on top of Scribe. The Scribe software
provides two methods, forward and deliver which are invoked by Pastry whenever a message
arrives, the former when the node is an intermediate stop towards the final destination, the latter
when the node is the final destination itself. The possible message types in Scribe are:
• Subscribe to a topic
• Unsubscribe from a topic
• Create a topic
• Publish an event to a topic
Each topic has a unique topic ID. The Scribe node with node ID numerically closest to the topic
ID acts as the rendezvous point (RP) for that topic. The RV point forms the root of the multicast
tree associated to the topic. To create a topic, a Scribe node asks Pastry to route a create message
to the RV point of the topic. The Scribe deliver method adds the topic to the set of topics that this
node is aware of. The topic ID can be the hash value of the topic’s textual name concatenated
with its creator’s name. The hash is computed using a collision resistant hash function (e.g. SHA-
1), which ensures a uniform distribution of topic IDs. Combining this with the uniform
distribution of Pastry’s node IDs, this ensures an even distribution of topics across nodes.
Figure 3. Pastry routing in a circular namespace (each dot depicts a live node in the namespace) [9]
Scribe creates a multicast tree per topic rooted at the RV point of the group to disseminate events
International Journal of Computer Networks & Communications (IJCNC) Vol.6, No.3, May 2014
19
along the tree. The multicast tree is created using reverse path forwarding, since the shortest paths
in the overlay towards the RV point are merged. Each node in the multicast tree is a forwarder for
the group. A forwarder may or may not be a subscriber to the group. Each forwarder maintains a
children table that contains an entry (IP address and node ID) for each of its children in the
routing table.
6) NICE [2] is an ALM protocol that arranges the set of members in a multicast group into a
hierarchical control topology. As new members join and existing members leave the group, the
basic operation of the protocol is to create and maintain the multicast tree hierarchy. The NICE
hierarchy is created by assigning members to different levels (or layers) as illustrated in layer 0,
Figure 4. Layers are numbered sequentially with the lowest layer of the hierarchy being layer zero
(L
0
). Members in each layer are partitioned into a set of clusters [13]. Each cluster is of size
between k and 3k-1 members, where k is a constant (usually k=3), and consists of a set of
members that are close to each other. Further, each cluster has a cluster leader. The protocol
chooses the center of the cluster to be its leader, i.e., the cluster leader has the minimum distance
to all other members in the cluster. This choice of the cluster leader ensures that a new joining
member is quickly able to find its appropriate position in the hierarchy using a very small number
of queries to other members. The leaders in level i are the members of level i+1 in the tree, so all
the leaders in L
0
belong to L
1
and their leaders belong to L
2
and so on until there is only one leader
which is the Rendezvous Point (RP) in the highest level of the tree. Since each cluster in the
hierarchy has between k and 3k -1 members, a host that belongs only to L
0
layer peers with O(k)
other hosts for exchange of control messages. In general, a host that belongs to layer L
i
and no
other higher layer, peers with O(k) other hosts in each of the layers L
0
....L
i
, which results in
control overhead for this member. Hence, the cluster-leader of the highest layer cluster
peers with a total of neighbors, which is the worst case control overhead at a
member. NICE mainly focuses on minimizing end-to-end delay. This is done by computing the
distance between the nodes and constructing the tree such that nodes close to each other get
assigned to the same cluster. This technique minimizes end-to-end delays as the cluster leader is
always centered in the middle of the cluster where the distance between it and the rest of the
cluster members is minimum.
Figure 4. Hierarchical arrangement of hosts in NICE [14]
3.
PROPOSED LINK-AWARE NICE (LA-NICE)
LA-NICE is an enhancement to the NICE protocol. NICE mainly focuses on minimizing end-to-
end delay. This is done by computing the distance between the nodes and constructing the tree
such that nodes close to each other get assigned to the same cluster. This technique minimizes
end-to-end delays as the cluster leader is always centered in the middle of the cluster where the
distance between it and the rest of the cluster members is minimum. LA-NICE takes into account
the fact that different links can have different bandwidths and uses this fact to improve multicast
message delivery and minimize end-to-end delay. As shown in Figure 5, LA-NICE doesn’t
International Journal of Computer Networks & Communications (IJCNC) Vol.6, No.3, May 2014
20
change the cluster structure or how the cluster splits or merges; it focuses on two phases in the
tree management which are the member join and the tree maintenance phases.
Figure 5. LA-NICE code structure
3.1. Multicast Tree Member Join Procedure
When a new node wants to join the multicast tree, in the original NICE algorithm, it contacts the
RP with its join query. The RP responds with the hosts that are present in the current highest level
of the tree L
i
. The joining host then contacts all members in L
i
to locate the member closest to it.
This member then informs the joining host of its other members in L
i-1
layer and so on recursively
until the joining host finds its position in the lowest level of the tree. Meanwhile, the joining host
gets peered temporarily with the RP as soon as it starts communicating with it to get the multicast
messages sent until it finds its appropriate position in the tree. This ensures that the joining node
gets connected to the tree as soon as it requests to join. Right before joining the tree in the lowest
level and after knowing where exactly it will join, the member node requests to be disconnected
from the RP and requests to join its appropriate parent in the tree.
LA-NICE modifies the member join procedure of NICE. NICE members join the closest clusters
based on round trip time (RTT) measurements. The RTT of sending a message is the time it takes
for the message to be sent plus the time it takes for an acknowledgment of that message to be
received. The RP gets a list of the clusters leaders ordered by distance and assigns the new host to
the closest cluster to it. In LA-NICE, in a tree of i levels, the RP gets a set of potential clusters
(PC) that the new node can join using Eq. (1)
(1)
Where PC is a set of clusters of size i. Then the RP checks the bandwidth of the leaders of the
potential clusters and assigns the new node to the cluster with the highest ratio of leader
bandwidth/number of cluster members as shown in Eq. (2).
(2)
3.2. Multicast Group Member Leave Procedure
When nodes leave the tree, they can leave either gracefully or ungracefully. A graceful leave,
which is when a host leaves the multicast group after notifying all the clusters it has previously
joined that it’s leaving. On the other hand, an ungraceful leave occurs when member fails without
being able to send out a leave notification to its clusters, the cluster members then manage to
International Journal of Computer Networks & Communications (IJCNC) Vol.6, No.3, May 2014
21
detect this departure when they don’t receive HeartBeat message from that member. A HeartBeat
message is a periodic message sent by every member to the rest of the multicast group informing
them that it still exists in the group. If a cluster leader left the group, this could lead to a partition
in the tree so a new leader needs to be chosen faster. A new leader of the cluster is chosen
depending on who is estimated to be closest to the center among these members.
3.3. Multicast Tree Maintenance
A cluster leader periodically checks the size of its cluster, and appropriately splits or merges the
cluster when it detects a size bound violation. If a cluster exceeds the cluster size upper bound 3k-
1, it gets split into two equal-sized clusters. Given a set of hosts and the distances between them,
the cluster split operation partitions them into subsets that meet the size bounds, such that the
maximum radius of the new set of clusters is minimized. The centers of the two partitions are
chosen to be the leaders of the new clusters and transfers leadership to the new leaders. If these
new clusters still violate the size upper bound, they are split by the new leaders using identical
operations.
To maintain the tree structure and detect any unexpected partitioning in the tree, each member of
a cluster sends periodic HeartBeat message to each of its cluster members. The message contains
the distance estimate from the sender to each other member of the cluster. The cluster leader
includes the complete updated cluster membership in its HeartBeat messages to all other
members. This allows existing members to set up appropriate peer relationships with new cluster
members on the control path. The cluster leaders also periodically send their higher layer cluster
membership their cluster.
LA-NICE takes the link load into account when maintaining the tree. So instead of only checking
if the cluster leaders need to be modified based on the leader’s position with respect to the cluster
members, LA-NICE checks the bandwidth of the closest 3 nodes to the center of the cluster
(given that the cluster has more than 3 members) and assigns the one with the highest ratio of
(bandwidth/number of cluster members) to be the cluster leader. This modification is selecting the
cluster leader based on the fact that there is always higher load on the cluster leaders than the
other nodes in the cluster since the cluster leaders send the multicast messages to all the cluster
members leading to a bottleneck of O(k log N). So, ensuring that the leader has a relatively high
bandwidth in addition to being close to all the members reduces the delay and improves multicast
message deliver.
4.
RESULTS AND ANALYSIS
4.1. Experiment Setup
To evaluate LA-NICE, we compared it to NICE protocol and Scribe [9] which is another ALM
protocol. The evaluation was done using four different test groups of users. Each group had
different number of users. The number of users in those groups were 20, 45, 70, and 100
respectively. The simulation was done using OMNeT++ as shown in Figure 6 and it runs for 300
seconds where nodes exchange multicast messages where some users would drop out while other
join the group randomly.
International Journal of Computer Networks & Communications (IJCNC) Vol.6, No.3, May 2014
22
4.2. OMNeT++
OMNeT++ [15] is an object-oriented open-source network simulator that is highly flexible and
easy to use. The model’s structure is described in OMNeT++ Network Description (NED)
language. NED facilitates the modular description of a network, which consists of a number of
component descriptions (channels, simple/compound module types, gates, etc.). In addition,
OMNeT++ provides various statistic collections and visualization tools for results analysis. The
simulator supports parallel and distributed simulation with the multiple instances communicating
via Message Passing Interface (MPI), as well as support for network emulation through interfaces
with real networks and the ability to use real networking code inside the simulator. OMNeT++
simulation models are composed of modules and connections. Connections may have associated
channel objects. Channel objects encapsulate channel behavior: propagation and transmission
time modeling, error modeling, and possibly others. Channels are also programmable in C++ by
the user. Modules and channels are called components. Components are represented with the C++
class cComponent.
4.3. Implementation
The implementation was done using OMNeT++ framework with oversim [16]. Oversim includes
a basic implementation of NICE protocol. They were both used in evaluating LA-NICE
performance. The implementation of LA-NICE was built using OMNeT++ oversim framework.
In LA-NICE, both the member join and the maintenance methods were modified. The code is
implemented in C++ to write the logic of the protocol and OMNeT’s.Net language to represent
the user interface of the network. Datarate channels were used with different datarates to test
various network conditions and bandwidth variations.
Figure 6. LA-NICE multicast tree simulation using OMNeT++
4.4. Results
Message Delivery: The first evaluation criteria for LA-NICE is the percentage of multicast
messages delivered, failure in delivery indicates inefficient tree maintenance due to not detecting
node departures in a reasonable amount of time or bad bandwidth utilization, which results in
bottlenecks that lead to late delivery and sometimes failure in delivery.
International Journal of Computer Networks & Communications (IJCNC) Vol.6, No.3, May 2014
23
Table 1. Message Delivery Percentage Results
Number
of users NICE
Scribe
LA
-
NICE
%
Improvement
over
NICE
over
Scribe
20 90.93
88.57
99.34
8.4%
10.7%
45 96.01
89.65
98.69
2.7%
9.04%
70 90.48
88.71
96.53
6.05%
7.8%
100 93.76
89.62
95.68
1.92%
6.06%
Figure 7. Percentage of Message Delivery of LA-NICE, NICE and Scribe against different number of users
As seen in Table 1 and Figure 7, scribe has lower message delivery percentage. The reason
behind the less performance in Scribe is due to the fact that the nodes are assigned random
generated IDs. This randomness could lead to a situation where two hosts can be close to each
other and yet sending a message from one of them to the other passes by other nodes that are far
from them which takes longer time than it should. This is due to the routing table which sends
messages to hosts that have the same prefix in their ID (which doesn’t always mean that this is the
closest node in the table). In addition, messages have a certain time to live (TTL) and then they
timeout, so as the delay increases the amount of messages that timeout increase. NICE, on the
other hand doesn’t have this problem and it sends the messages to the closest nodes graphically
without any regard to bandwidth conditions. LA-NICE combines both distance and bandwidth
factors.
On the other hand, NICE takes the node’s proximity to each other into account when constructing
the delivery tree. For example, if two nodes are close to each other, the time taken to send
exchange messages between them should be less than the time taken to exchange message
between nodes that are further away from each other.
Our main focus is to build on NICE and enhance its performance by making it link-aware. Link-
Aware NICE handles the proximity factor in NICE in addition to the bandwidth factor when a
new node joins the tree and when maintaining the tree as well. This is done through measuring the
cluster leaders’ bandwidth and selecting the cluster with the highest leader bandwidth that is
relatively close. This approach produced the highest message delivery percentage as seen in
International Journal of Computer Networks & Communications (IJCNC) Vol.6, No.3, May 2014
24
Figure 7 compared to NICE and Scribe.
Table 2 Average Hop Count Results
Number of
users LA-NICE NICE
20
1.55
1.48
45
1.73
1.77
70
1.79
2.1
100
1.78
2
Table 3. Maximum Hop Count Results
Number of
users LA-NICE NICE
20
2
2
45
4
4
70
4
7
100
3
6
Message Delay and Hop count: Another evaluation criteria is the average hop count of the
multicast messages. The hop count of a packet is defined as the number of routers traversed by a
packet between its source and destination [17], which is in this case the number of hosts that a
message passes from the source to the destination [18]. As seen in Figures 8 and 9 and Tables 2
and 3 when the number of users is small, the average hop count of LA-NICE and NICE, as well
as the maximum hop count is almost the same, however as the number of users increase LA-
NICE has less number of hops leading to less delay in delivering the messages.
Figure 8. Average hop count in LA-NICE and NICE against different number of users
International Journal of Computer Networks & Communications (IJCNC) Vol.6, No.3, May 2014
25
Figure 9. Maximum hop count in LA-NICE and NICE against different number of users
Table 4. Average Delay Results (S)
Number of users LA-NICE NICE
20
0.923
0.91
45
0.83
0.9
70
1.54
1.58
100
1.21
1.19
Table 5. Maximum Delay Results (S)
Number of users LA-NICE NICE
20
1.67
1.72
45
1.76
1.57
70
2.48
2.48
100
2.01
2.05
The delay factor is related to the hop count. Figures 10 and 11 and Tables 4 and 5 show a
comparison between LA-NICE, NICE in terms of delay. The delay and message hops are usually
proportional, therefore, with the increase of the hops along the tree, the delay increases. It is clear
that LA-NICE outperforms NICE in terms of delay and hop count after taking the links’ load
factor into account.
International Journal of Computer Networks & Communications (IJCNC) Vol.6, No.3, May 2014
26
Figure 10. Average Delay in LA-NICE and NICE against different number of users
Figure 11. Maximum Delay in LA-NICE and NICE against different number of users
R
EFERENCES
[1] D. Helal, A. Naser, M. Rehan, and A. ElNaggar, Link-aware nice application level multicast protocol,
The Ninth International Conference on Digital Telecommunications (2014) 8–12.
[2] S. Banerjee, B. Bhattacharjee, and C. Kommareddy, Scalable applica-tion layer multicast, vol. 32, no.
4. New York, (2002) 205–217.
[3] A. Varga and R. Hornig, An overview of the omnet++ simulation environment, 1st International
Conference on Simulation Tools and Techniques for Communications, Institute for Computer
Sciences, Social-Informatics and Telecommunications Engineering, Brussels, Belgium (2008) 60:1–
60:10.
[4] K. Ke and C. Huang, Performance evaluation of multisource applica-tion layer multicast (alm):
Theoretical and simulative aspects, vol. 57, no. 6. New York (2013) 1408–1424.
[5] L. Lao, J.-H. Cui, M. Gerla, and D. Maggiorini, A comparative study of multicast protocols: top,
bottom, or in the middle?, INFOCOM 2005. 24th Annual Joint Conference of the IEEE Computer and
Communications Societies. (2005) 2809–2814.
[6] B. Rong and A. Hafid, A distributed relay selection algorithm for cooperative multicast in wireless
mesh networks. 5th International Conference in Mobile Ad-hoc and Sensor Networks (2009) 49–56.
International Journal of Computer Networks & Communications (IJCNC) Vol.6, No.3, May 2014
27
[7] V. N. Padmanabhan, H. J. Wang, P. A. Chou, and K. Sripanidkulchai, Distributing streaming media
content using cooperative networking. NOSSDAV. ACM, (2002) 177–186.
[8] D. A. Tran and et al., Zigzag: An efficient peer-to-peer scheme for media streaming, IN PROC. OF
IEEE INFOCOM, 2003.
[9] M. Castro, P. Druschel, A. Kermarrec, and A. Rowstron, Scribe: A large-scale and decentralized
application-level multicast infrastructure, vol. 20 (2002) 1489 – 1499.
[10] M. Hosseini, D. T. Ahmed, S. Shirmohammadi, and N. D. Georganas, A survey of application-layer
multicast protocols, IEEE Commun. Surveys and Tutiruals (2007).
[11] Y. Chu, S. G. Rao, S. Seshan, and H. Zhang, A case for end system multicast, ACM Sigmetrics,
(2000) 1–12.
[12] A. Rowstron, A. Kermarrec, M. Castro, and P. Druschel, Scribe: The design of a large-scale event
notification infrastructure, Networked Group Communication, 2001, pp. 30–43.
[13] X. Li, X. Zhang, W. Luo, and B. Yan, A clustering scheme in application layer multicast. vol. 30, no.
2, (2011) 335–355.
[14] W. Xijuan, J. Ruisheng, L. Guang, and Y. Xianghong, Research on p2p-based application layer
multicast technology for streaming media, IEEE Computer Society, Los Alamitos, (2010) 341–345.
[15] A. Varga, Omnet++ discrete event simulation system user manual (2011).
[16] I. Baumgart, B. Heep, and S. Krause, Oversim: A scalable and flexible overlay framework for
simulation and real network applications, Peer-to-Peer computing. IEEE Ninth International
Conference on P2P, (2009) 87–88.
[17] C. Hubsch,¨ C. P. Mayer, and O. P. Waldhorst, User-perceived performance of the nice application
layer multicast protocol in large and highly dynamic groups, 15th international GI/ITG conference on
Measurement, Modelling, and Evaluation of Computing Systems and Dependability and Fault
Tolerance. Berlin (2010) 62–77.
[18] S. Rizvi, M. Khan, and A. Riasat, Deterministic formulization of bandwidth efficiency for multicast
systems, 2nd International Conference on Computer, Control and Communication (2009) 1–6