An Adaptive Load Sharing Algorithm for Heterogeneous Distributed System

International Journal of Research in Computer Science
eISSN 2249-8265 Volume 3 Issue 3 (2013) pp. 9-15
www.ijorcs.org, A Unit of White Globe Publications
doi: 10.7815/ijorcs. 33.2013.063
www.ijorcs.org
AN ADAPTIVE LOAD SHARING ALGORITHM FOR
HETEROGENEOUS DISTRIBUTED SYSTEM
P. Neelakantan
Department of CSE, SVUCE, Tirupati, A.P, INDIA
E-mail: pneelakantan@rediffmail.com
Abstract: Due to the restriction of designing faster and
faster computers, one has to find the ways to maximize
the performance of the available hardware. A
distributed system consists of several autonomous
nodes, where some nodes are busy with processing,
while some nodes are idle without any processing. To
make better utilization of the hardware, the tasks or
load of the overloaded node will be sent to the under
loaded node that has less processing weight to
minimize the response time of the tasks. Load
balancing is a tool used effectively for balancing the
load among the systems. Dynamic load balancing
takes into account of the current system state for
migration of the tasks from heavily loaded nodes to the
lightly loaded nodes. In this paper, we devised an
adaptive load-sharing algorithm to balance the load
by taking into consideration of connectivity among the
nodes, processing capacity of each node and link
capacity.
Keywords: Load balancing, Distributed System,
heterogeneous, response time.
I. INTRODUCTION
An important attribute in a dynamic load balancing
policy is to initiate the load balancing activity. The
balancing activity specifies which node is responsible
for detecting imbalance of the load among the
nodes[9]. A load-balancing algorithm is invoked when
load imbalance among the nodes is detected. The
initiation of load balancing activity will have a higher
impact on complexity, overhead and scalability. The
load balancing algorithm is designed in such a way to
make the overloaded node to transfer its excess load to
the underloaded node which is called sender – initiated
and when underloaded node requests the load from the
overloaded node then it is called receiver- initiated [6]
[8].
Domain balancing is used to decentralize the
balancing process by minimizing its scope and
decrease the complexity of the load balancing
algorithm. A domain is defined as subset of nodes in a
system, such that a load balancing algorithm can be
applied for this subset of nodes in a single step.
Domain balancing is used in load balancing algorithms
to decentralize the balancing. The balancing domains
are further divided into two types: The first type is
overlapped domains, which consists of node initiating
the balancing activity and balances its load by
migrating the tasks or load units with the set of
surrounding nodes [3].
Global balancing is achieved by balancing every
domain and by diffusing the excess load throughout
the overlapped domains in a distributed system.
Another important attribute in load balancing
algorithm is the degree of information. The degree of
information plays an important role in making the load
balancing decisions. To achieve global load balancing
in a few steps, the load balancing should get absolute
information instead of getting the obsolete information
from the nodes. In general, the collection of
information by a node is restricted to the domain or
nearest neighboring nodes (which are directly
connected to a node) [4].
Although collecting information from all the nodes
in a distributed system gives the exact knowledge of
the system, it introduces large communication delay,
so from this perspective, it will have a negative impact
on the load balancing algorithm. In such cases, it has
been observed, that the average response time is kept
minimum without load balancing instead of doing the
load balancing which induces overhead in migrating
the load from one node to another node in the system
[5].
In this section, an abstract view of the software
details is presented for load balancing. The distributed
system consists of several nodes and the same load
balancing software is installed to run on all the nodes
in the distributed system. By installing the same
software in all the nodes, the load balancing decision is
taken by a node locally (decentralized) by collecting
the information from the neighboring nodes as
opposed to the centralized load balancing policy[14].
The program must use a multi-threaded concept to
implement load balancing in a distributed system. Two
communication ports are available: TCP and UDP.
UDP is preferable as it incurs less communication
overhead. In general the architecture provides three

10 P. Neelakantan
www.ijorcs.org
layers: Communication layer, Load balancing process
and application layer[14][10]. For storing information
two data structures were used.
The communication link is responsible for four
phases: node status information phase, node status
reception, tasks reception and task migration. The node
status information is responsible for disseminating the
load information to the node that has requested it. The
exchange of the information has a profound effect on
the load balancing decision; it has to be done
according to the predefined intervals of time specified
on each node [7] [14].
The status reception is responsible for receiving the
status information from the other nodes and it will be
updated in the local node list which is running the
status reception phase. Here it is possible to distinguish
the old information from the new information. The
technique that is used to find is to associate the
timestamp for the information that it has received from
some node (say 𝑇𝑆𝑗
𝑖
(𝐼𝑛𝑓), the time stamp attached to
the information received from 𝑗 to 𝑖). The local node
say 𝑖 maintaining the status about the node 𝑗 is kept in
the memory. If any estimate regarding node 𝑗 exists in
the node 𝑖 memory, it will be compared to the received
time stamp message and drops the old time stamp and
the new timestamp message has been saved in the
memory as the old time stamp has the obsolete
information[11][1][2].
Once a node collects the above information, then it
knows whether it is overloaded or underloaded. In case
if it is overloaded node, then it transmits the excess
tasks (loads) to the underloaded nodes in a “tasks
transmission” phase. The next initiation of load
balancing activity will be done only when the current
migration of load units to the underloaded nodes is
completed.
The “task reception” is responsible for listening to
the requests and accepts the tasks sent from the other
nodes. As we can observe from the above situations,
the minimum time to initiate the new load balancing
activity takes three time instants. One instant for
receiving the status of all the nodes and second time
instant for determining the underloaded nodes and
computing the excess load and third time instant for
transferring the excess load to the underloaded nodes
which has been determined in the second time instant.
So, the new load balancing activity takes place only at
the fourth time instant [12] [14].
In a few papers [3] [9] [10], it is assumed that the
nodes will not fail. The problem arises when the nodes
fail which is common in the distributed systems.
Sometimes a communication link will also fail, so the
node will be unreachable. These two aspects i.e.,
failure of a node and the communication link will
affect greatly the load balancing algorithms. Let us
assume the following scenario. The overloaded node
has collected the load information from the
neighboring nodes and found some of the nodes are
low loaded as discussed earlier. Now at the given time
instant when the node tries to send its excess load to
the overloaded node, it will not succeed because of the
failure of the node. The node may fail after sending the
status information. If this happens, an alternative must
be chosen to avoid a failure of the load-balancing
algorithm.
II. NOTATIONS & ASSUMPTIONS
N : Number of nodes
V= {1, 2,…, N} a set of nodes in a system
xi(t) : Expected waiting time experienced by a task
inserted into the queue at the ith
node in time t
Ai(t) : rate of generation of waiting time on ith
node
caused by the addition of tasks in time t.
Si(t) : rate of reduction in waiting time caused by the
service of the tasks at the ith
node in time t.
ri(t) : rate of removal(transfer) of the tasks from node j
to node i at time t by the load balancing algorithm at
node j.
tsi : Average completion time of the task at node i.
bi : Average size of the task in bytes at node i when it
is transferred
dij : Transfer rate in bytes/sec between node i and node
j
𝑥̅ 𝑖(𝑡): Average size of the queue calculated by node i
based on its domain information at time t.
𝐷𝑖: Neighboring nodes to i which is defined as
𝐷𝑖 = {𝑗|𝑗 ∈ 𝑉 𝑎𝑛𝑑 (𝑖, 𝑗) ∈ 𝐸} where V= {1, 2…N}
𝐸𝑖(𝑡): Excess number of tasks at node i at time t.
fij : Portion of the excess tasks of node i to be
transferred to node j decided by the load balancing
algorithm.
The following assumptions were made in this paper:
1. It is assumed that a distributed system consists of N
heterogeneous nodes interconnected by an
underlying arbitrary communication network. Each
node i in a system has a processing weight Pi >0
and processing capacity Si>0. The load is defined to
be Li= Pi/Si. In homogenous case the value of
Li=Pi.
2. It has been assumed that tasks arrive at node i
according to Poisson process with rate 𝜆𝑖(𝑡). A task
arrived at node i may be processed locally or
migrated through the network to another node j for
remote processing. Once the task is migrated, it
remains there until its completion.

An Adaptive Load Sharing Algorithm for Heterogeneous Distributed System 11
www.ijorcs.org
3. It is assumed that there is a communication delay
incurred when task is transferred from one node to
another before the task can be processed in the
system. The communication delays are different for
each link.
Each node contains an independent queue where
arrived tasks are added to the queue, which results in
accumulation of waiting time. Load balancing must be
done repeatedly to maintain load balance in the
system. The proposed algorithm is distributed in nature
meaning that each node runs load-balancing algorithm
autonomously.
The second level of the system is a load-balancing
layer, which consists of load balancing algorithms. The
load balancing process is initiated by suing predefined
time instants or randomly generating which is kept in a
file. The algorithm determines the portion of the
excess load to be sent to the underloaded node based
on the current state of the node and availability of the
nodes in the network. The load balancing algorithm
must consider the communication delay while
migrating the tasks to the other nodes. The algorithm
selects the tasks to migrate to other nodes by setting
their status as inactive to avoid execution of the tasks
by current node application during the transition
period. After completion of the task transmission
activity, the status of the tasks is set to active when
they are not transmitted to any node. when the tasks
are transmitted to other nodes during the task
transmission phase then those tasks are removed from
the task queue of the current node.
Application layer consists of two threads of control:
Task input and task execution threads. The task input
creates a number of tasks defined in the initialization
file and inserts them in the task queue. This task input
is also responsible for adding the new tasks to the task
queue either from the current node or from other nodes
in the system. The task execution thread is responsible
for execution of the tasks and updating the QSize
variable by removing the task from the task queue.
The load balancing policy must take into account of
processing capacity of the node while migrating the
tasks to it. The selected node may become a candidate
for one or more overloaded node in a given time
instant because of the decentralized policy. Another
issue to be considered is variable task completion
times. Taking these issues in priori is not possible so a
load balancing strategy must be adaptive to the
dynamic state changes in the system and act
accordingly to transfer the tasks. Even this can result
in task shuttle between the nodes, so a migration limit
for a task should be set to avoid task thrashing.
Another issue to be considered while migrating the
tasks from one node to another node in a system is
communication overhead. Large communication
delays will have a negative impact on the load
balancing policy, so, the transfer delays must be taken
into account while migrating the task. When the
completion of the task time in current node is greater
than the completion time on task in another node
inclusive of communication overhead, then only a task
is considered for migration.
III. MATHEMATICAL MODEL
The mathematical model for load balancing in a
given node i is given by [1][2]
𝑑𝑤 𝑖(𝑡)
𝑑𝑡
=𝐴𝑖 − 𝑆𝑖 + 𝑟𝑖(𝑡) − ∑ 𝑓𝑖𝑗
𝑡𝑠𝑖
𝑡𝑠 𝑗
≠𝑁 𝑖
𝑗=1 𝑟𝑗(𝑡 − 𝜏𝑖𝑗) (1)
𝐸𝑖(𝑡)= 𝑞𝑖(𝑡)- 𝑞�𝑖(𝑡)
𝑟𝑖(𝑡) = 𝐺𝑖(𝐸𝑖(𝑡))
𝑓𝑖𝑗 ≥ 0, 𝑓𝑖𝑖=0, ∑ 𝑓𝑖𝑗 = 1
≠𝑁𝑖
𝑗=1
𝐸𝑖(𝑡) = �
𝐸 𝑖𝑓 𝑦 ≥ 0
0 𝑖𝑓 𝑦 < 0
When a task is inserted into the task queue of node
i, then it experiences the expected waiting time which
is denoted by wi(t).
Let the number of tasks in ith
node is denoted by 𝑞𝑖(t).
Let the average time needed to service the task at node
i 𝑡𝑠𝑖 .
The expected (average) waiting time is given by at
node i is given by 𝑤𝑖(𝑡) = 𝑞𝑖(𝑡)𝑡𝑠𝑖.
Note that 𝑤𝑖(𝑡)/𝑡𝑠𝑖 = 𝑞𝑖 is the number of tasks in
the node i queue.
Similarly 𝑤 𝑘(𝑡)/𝑡𝑠 𝑘 = 𝑞 𝑘 is the queue length of
some node k. If tasks on node i were transferred to
some node k, then the waiting time transferred is
𝑞𝑖 𝑡𝑠 𝑘=
𝑤 𝑖(𝑡)𝑡𝑠 𝑘
𝑡𝑠𝑖
, so that the fraction 𝑡𝑠 𝑘/𝑡𝑠𝑖 converts
waiting time on node i to waiting time on node k.
𝐴𝑖 ∶ Waiting time generated by adding the task in the
ith
node.
𝑆𝑖 : Rate of reduction in waiting time caused by the
service of tasks at the ith
node is given by 𝑆𝑖 = (1 ∗
𝑡𝑝𝑖)/𝑡𝑝𝑖=1 for all 𝑤𝑖(𝑡) > 0.
𝑟𝑖(𝑡) ∶ The rate of removal (transfer) of the tasks
from node i at time t by the load balancing algorithm at
node i. 𝑓𝑖𝑗 is the fraction of ith
node tasks to be sent out
to the jth
node. In more detail fijri(t) is the rate at which
node i sends waiting time (tasks) to node i at time t
where fii>=0 and 𝑓𝑖𝑖=0.That is, the transfer from node i
of expected waiting time (tasks) ∫ 𝐸𝑖(𝑡)𝑑𝑡
𝑡2
𝑡1
in the
interval of time [𝑡1, 𝑡2] to the other nodes is carried out
with the 𝑗 𝑡ℎ
node receiving the fraction 𝑝𝑖𝑗(𝑡 𝑝 𝑗
/
𝑡 𝑝𝑖
) ∫ 𝑢𝑖(𝑡)𝑑𝑡
𝑡2
𝑡1
where the ratio 𝑡 𝑝 𝑗
/𝑡 𝑝𝑖
converts the
task from waiting time on node i to waiting time on

12 P. Neelakantan
www.ijorcs.org
node j. As ∑ (𝑛
𝑖=1 𝑓𝑖𝑗 ∫ 𝐸𝑖(𝑡)𝑑𝑡
𝑡2
𝑡1
) = ∫ 𝐸𝑖(𝑡)𝑑𝑡
𝑡2
𝑡1
, this
results in removing all of the waiting time ∫ 𝐸𝑖(𝑡)𝑑𝑡
𝑡2
𝑡1
from node i.The quantity 𝑓𝑖𝑗 𝐸𝑖(𝑡 − 𝜏𝑖𝑗) is the rate of
increase (rate of transfer) of the expected waiting time
(tasks) at time t from node i by (to) node j where
𝜏𝑖𝑗(𝜏𝑖𝑖 = 0) is the time delay for the task transfer from
node i to node j.
In this model, all rates are in units of the rate of
change of expected waiting time, or time/time which is
dimensionless. As 𝐸𝑖(𝑡) ≥ 0, node i can only send
tasks to other nodes and cannot initiate transfers from
another node to itself. A delay is experienced by
transmitted tasks before they are received at the other
node. The control law 𝐸𝑖(𝑡) = 𝐺𝑖 ∗ 𝐸𝑖(𝑡) states that if
the 𝑖 𝑡ℎ
node output 𝑤𝑖(𝑡) is above the domain average
(∑ 𝑞 𝑗(𝑡 − 𝜏𝑖𝑗))𝑛
𝑗=1 /n, then it sends data to the other
nodes, while if it is less than the domain average
nothing is sent. The 𝑗 𝑡ℎ
node receives the fraction
∫ 𝐹𝑖𝑗
𝑡2
𝑡1
(𝑡 𝑝𝑖
/𝑡 𝑝 𝑗
) 𝑢𝑖(𝑡)𝑑𝑡 of transferred waiting time
∫ 𝐸𝑖(𝑡)𝑑𝑡
𝑡2
𝑡1
delayed by the time 𝜏𝑖𝑗.The model
described in (1) is the basic model for load balancing,
but an important feature is to determine fij for each
underloaded node j. One approach is to distribute the
excess load equally to all the underloaded neighbors.
𝑓𝑖𝑗 =
1
𝑛−1
for i≠j.
Another approach is to use the load information
collected from the neighbors to determine the deficit
load of the neighbors. The deficit load of the neighbors
shall be determined by node i by using the formula (2)
𝑞 𝑗(𝑡-𝜏𝑖𝑗) − 𝑞�𝑖 (2)
The above formula is used by node i to compute the
deficiency waiting times in the queue of node j with
respect to the domain load average of node i.
If node j queue is above the domain average
waiting time, then node i do not send tasks to it.
Therefore (𝑞�𝑖 − 𝑞 𝑗(𝑡-𝜏𝑖𝑗)) is a measure by node i as
how much node j is behind the domain average waiting
time. Node i performs this computation for all the
other nodes which are directly connected to it and then
portions out its tasks among the other nodes that fall
below the domain queue average of node i.
𝑓𝑖𝑗 =
(𝑞� 𝑖−𝑞 𝑗(𝑡−𝜏 𝑖𝑗))
∑ (𝑞� 𝑖−𝑞 𝑗(𝑡−𝜏 𝑖𝑗)
𝑁 𝑖
𝑗=1
(3)
If the denominator ∑ (𝑞�𝑖 − 𝑞 𝑗(𝑡 − 𝜏𝑖𝑗)
𝑁 𝑖
𝑗=1 =0 then fij
are defined to be zero then no waiting times are
transferred. If the denominator ∑ (𝑞�𝑖 − 𝑞 𝑗(𝑡 −
𝑁 𝑖
𝑗=1
𝜏𝑖𝑗)=0, then(𝑞�𝑖 − 𝑞 𝑗(𝑡 − 𝜏𝑖𝑗) ≤ 0∀𝑗 ∈ 𝑁𝑖.
However by definition of the average ∑ (𝑞�𝑖 −
𝑁 𝑖
𝑗=1
𝑞 𝑗(𝑡 − 𝜏𝑖𝑗)+𝑞�𝑖 − 𝑞𝑖(𝑡) =∑ (𝑞�𝑖 − 𝑞 𝑗(𝑡 − 𝜏𝑖𝑗)
𝑁 𝑖
𝑗=1 )=0
which implies
𝑞�𝑖 − 𝑞 𝑗(𝑡)=∑ (𝑞�𝑖 − 𝑞 𝑗(𝑡 − 𝜏𝑖𝑗)
𝑁 𝑖
𝑗=1 ) > 0
That is, if the denominator is zero, the node j is not
greater than its domain queue average, so Ei(t)= Gi
.Ei(t))=0, where G is Gain Factor.fij : Portion of the
excess tasks of node i to be transferred to node j
decided by the load balancing algorithm.
Except the last three parameters remaining
information is known at the time of load balancing
process. Before the instance of load balancing activity,
every variable is updated.
IV. ALGORITHM
A. Algorithm ALS
The current node i, performs the followings:
a. Calculate the average queue size ( 𝑞�𝑖)based on the
information received from the neighbouring nodes.
𝑞�𝑖=
1
≠𝑁𝑖+1
∑ (𝑞𝑖 + 𝑞 𝑗
𝑡𝑠 𝑗
𝑡𝑠𝑖
≠𝑁𝑖
𝑗=1 )
if (𝑞𝑖 > 𝑞�𝑖)then Ei =(𝑞𝑖-𝑞�𝑖) * G
else Exit.
b. Determine the participant nodes in load sharing
process.
Participants= {j| 𝑞 𝑗<𝑞�𝑖, ∀j∈Ni}
c. Calculate the fraction of the load ( 𝑓𝑖𝑗
′
) to be sent to
the participants
𝑓𝑖𝑗
′
=
𝑞�𝑖−(
𝑡𝑠 𝑗
𝑡𝑠 𝑖
)𝑞 𝑗
∑ ( 𝑞�𝑖−(
𝑐 𝑗
𝑐 𝑖
)𝑞 𝑗
𝑁 𝑖
𝑗=1
d. Calculate maximum portion of the excess load
(𝑓𝑖𝑗
′′
)
𝑓𝑖𝑗
′′
=
(𝑞𝑖−Ei) 𝑡𝑠𝑖 dij
Eibi
e. 𝑓𝑖𝑗 = Min (𝑓𝑖𝑗
′
, 𝑓𝑖𝑗
′′
)
f. For j∈ Participants
i. Announce to node j about its willingness to
send Tij= 𝑓𝑖𝑗 *Ei tasks;
ii. nowReceived = call procedure acceptanceFrom
Nodej()
iii. if(nowReceived >0)
1. Transfer NowReceived to j
2. Tij=Tij- NowReceived
End if
g. Repeat steps from (a) to (f).
Procedure acceptanceFromNodej()
if ((𝑞 𝑗+ Tij)< 𝑞� 𝑗nowSend=𝑞� 𝑗 − 𝑞 𝑗;
else nowSend=-1;

www.ijorcs.org
return now Send;
end acceptanceFromNodej
In general it is assumed that keeping the Gain factor
G=1 will give the good performance. But in a
distributed system with largest delays and the nodes
that have domain queue average outdated gives poor
result. This phenomenon was first observed by the load
balancing group at the University of New Mexico [7].
So the G values are set in the way that yields an
optimal result. Another step that is added in the above
algorithm is to test the node availability. It checks both
node availability as well as the amount of waiting
times it can receive. The node executing the ALS is
permitted to send the tasks to the neighbors after
receiving the acknowledgement specifying the amount
of the load they can be able to process.. The time
complexity of the proposed algorithm is O(d) shown in
the table 1.
Table 1: ALS Operations
Sno Actions Operation
Quantity,
(d is the
number of
neighbors)
1
Compute
average
queue size
Addition
Division
Multiplication
d+1
d
d
2 Compute Ei
Subtraction
Multiplication
1
1
3
Determine
the
participant
nodes
Comparison d
4
Compute
𝑓𝑖𝑗
′
Subtraction
Division
Multiplication
d+1
d+1
d+1
5
Compute
𝑓𝑖𝑗
′′
Subtraction
Division
Multiplication
1
1
3
6 Compute Tij Multiplication d
7
Message to
node
Transfer d
8
Compute
nowReceived
Addition
Comparison
Message
Transfer
d
d
d
V. SIMULATION
To test the performance of the newly proposed
load-balancing policy, a Java program is developed to
test the performance of the existing and proposed
algorithms. The existing algorithms, ELISA and
DOLB are used to compare with the proposed
algorithm ALS. The DOLB is very much related to the
above problem. The initial settings and parameters are
shown in Table 2. The average network transfer rates
between each node are represented by the cost
adjacency matrix.
The proposed algorithm ALS is tested with DOLB
& ELISA for the gain values G between 0.3 and 1 with
0.1 incremental steps. The 𝛼 parameter introduced in
the previous section was set to 0.05 by running several
experiments and observing the behavior of the tsi
parameter. Note that, the first time the load-balancing
process was triggered was after 40s from the start of
the system and then the strategy was executed
regularly at 20s interval.
Table 2: Simulation Parameters
Number of nodes 16,32,64
Initial task
distribution
[100..1000] tasks distributed
randomly at each node
Average task
processing time(𝑡𝑠
in ms)
Processing time is randomly
distributed in a
range[300…800]
Size of task( in Mb) 1
Load balancing
instance
First time the load balancing
was triggered at 5s then for
every 10s the load balancing
is initiated
Bandwidth
distribution (𝑑𝑖𝑗)
A cost adjacency matrix
denotes the transfer rate
between the nodes.
Number of nodes 16,32,64
Initial task
distribution
[100..1000] tasks distributed
randomly at each node
Average task
processing time(𝑡𝑠
in ms)
Processing time is randomly
distributed in a
range[300…800]
This was done to ensure that the ts parameter had
enough time to adapt and reflect the current
computational power of each node before the
occurrence of any tasks migration between the nodes.
Note that the ratio
𝑡𝑠𝑖
𝑡𝑠 𝑗
are fixed over time. The
proposed and rival methods were evaluated by
conducting 10 runs for each value of G between 0.3
and 1 with 0.1 incremental step.

14 P. Neelakantan
www.ijorcs.org
Figure 1: Completion time averaged over 5 runs vs different
gain values K. The graphs shows the results of three policies
for system size=64.
Figure 2: Completion time averaged over 5 runs vs.
different gain values K. The graphs shows the results of
three policies for system size=32.
Figure 3: Total number of tasks exchanged averaged over 5
runs Vs different Gain values K. The graphs shows the
performance of the three policies for system size=16.
VI. CONCLUSION
The proposed algorithm is better when compared to
the existing algorithms in the literature. In simulation,
we assumed the tasks with no precedence and with no
deadlines. However, as a future work, the algorithm
must focus on considering the tasks with dead line and
tasks with precedence relations.
VII. REFERENCES
[1] M. M. Hayat, S . Dhakal, C. T. Abdallah I ‘Dynamic
time delay models for load balancing. Part II: Stochastic
analysis of the effect of delay uncertainty, CNRS-NSF
Workshop: Advances in Control of tirne-delay Systems,
Paris France, January 2003.
[2] J. Ghanem, C. T. Abdallah, M. M. Hayat, S. Dhakal,
J.D Birdwell, J. Chiasson, and Z. Tang. Implementation
of load balancing algorithms over a local area network
and the internet. 43rd IEEE Conference on Decision
and Control, Submitted, Bahamas, 2004. doi:
10.1109/CDC.2004.1429411
[3] L. Anand, D. Ghose, and V. Mani, “ELISA: An
Estimated Load Information Scheduling Algorithm for
Distributed Computing Systems,” Int’l J. Computers
and Math With Applications, vol. 37, no. 8, pp. 57-85,
Apr. 1999. doi: 10.1016/S0898-1221(99)00101-7
[4] WEI Wen-hong, XIANG Fei, WANG Wen-feng, et al.
Load Balancing Algorithm in Structure P2P
Systems[J],Computer Science, 2010, 37(4):82-85.
[5] Khalifa, A.S.; Fergany, T.A.; Ammar, R.A.; Tolba,
M.F,” Dynamic online Allocation of Independent tasks
onto heterogeneous computing systems to maximize
load balancing,” IEEE International Symposium on
Signal Processing and Information Technology, ISSPIT
2008,Page(s): 418 – 425. doi: 10.1109/ISSPIT.
2008.4775659
[6] Andras Veres and Miklos Boda. The chaotic nature of
TCP congestion control. In Proceedings of the IEEE
Infocom, pages 1715-1723, 2000. doi:
10.1109/INFCOM.2000.832571
[7] J. Chiasson, J. D. Birdwell, Z. Tang, and C.T. Abdallah.
The effect of time delays in the stability of load
balancing algorithms for parallel computations. IEEE
Conference on Decision and Control, Maui, Hawaii,
2003. doi: 10.1109/CDC.2003.1272626
[8] Ming wu and Xian-He sun, A General Self Adaptive
Task Scheduling System for Non Dedicated
Heterogeneous Computing, In Proceedings of IEEE
International Conference on Cluster Computing, PP
354-361, Dec 2003. doi: 10.1109/CLUSTR.
2003.1253334
[9] Z. Zeng and B. Veeravalli, "Design and Performance
Evaluation of Queue-and-Rate-Adjustment Dynamic
Load Balancing Policies for Distributed Networks",
presented at IEEE Trans. Computers, 2006, pp.1410-
1422. doi: 10.1109/TC.2006.180
[10] K. Lu, R. Subrata, and A. Y. Zomaya, Towards
decentralized load balancing in a computational grid
environment, in: Proceedings of the first International
Conference on Grid and Pervasive Computing, 2006,
Vol. 3947, pp. 466-477, Springer-Verlag Press. doi:
10.1007/11745693_46
[11] Acker, D., Kulkarni, S. 2007. A Dynamic Load
Dispersion Algorithm for Load Balancing in a
Heterogeneous Grid System. IEEE Sarnoff
0
20
40
60
80
100
120
140
160
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
ALS
DOLB
ELISA
0
20
40
60
80
100
120
140
160
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
AverageTaskCompletiontime
Gain
ALS
DOLB
ELISA
0
50
100
150
200
250
300
350
400
450
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
NUmberofTasksexchanged
ALS
DOLB
ELISA

www.ijorcs.org
Symposium, 1- 5, 2007. 10.1109/SARNOF.2007.
4567375
[12] M. Luczak and C. McDiarmid. On the maximum queue
length in the supermarket model. The Annals of
Probability, 34(2):493–527, 2006. doi: 10.1214/
00911790500000710
[13] Zhou, S. (1987). An Experimental Assessment of
Resource Queue Lengths as Load Indices. Proc. Winter
USENIX Conf., p.73-82.
[14] J. Ghanem, C. T. Abdallah, M. M. Hayat, S. Dhakal,
J.D Birdwell, J. Chiasson, and Z. Tang. Implementation
of load balancing algorithms over a local area network
and the internet. 43rd IEEE Conference on Decision
and Control, Submitted, Bahamas, pp 4199-4204, 2004.
How to cite
P. Neelakantan, " An Adaptive Load Sharing Algorithm for Heterogeneous Distributed System ". International
Journal of Research in Computer Science, 3 (3): pp. 9-15, May 2013. doi: 10.7815/ijorcs. 33.2013.063

An Adaptive Load Sharing Algorithm for Heterogeneous Distributed System

More Related Content

What's hot (20)

Viewers also liked (9)

Similar to An Adaptive Load Sharing Algorithm for Heterogeneous Distributed System (20)

More from IJORCS (20)

Recently uploaded (20)

An Adaptive Load Sharing Algorithm for Heterogeneous Distributed System