Load balancing in Distributed Systems

1
ABSTRACT
A number of load balancing algorithms were developed in order to improve the execution of
a distributed application in any kind of distributed architecture. Load balancing involves
assigning tasks to each processor and minimizing the execution time of the program. In
practice, it would be possible even to execute the applications on any machine of worldwide
distributed systems. However, the ‘distributed system’ becomes popular and attractive with
the introduction of the web. This results in a significant performance improvement for the
users. This paper describes the necessary, newly developed, principal concepts for several
load balancing techniques in a distributed computing environment. This paper also includes
various types of load balancing strategies, their merits, demerits and comparison depending
on certain parameters. Distributed computing is a promising technology that involves
coordinate and involvement of resources to carry out multifarious computational problems.
One of the major issues in distributed systems is to design of an efficient dynamic load
balancing algorithm that improves the overall performance of the distributed systems.
Scheduling and resource management plays a decisive role in achieving high utilization of
resources in grid computing environments. Load balancing is the process of distributing the
load among various nodes of a distributed system to improve both job response
time and resource utilization while also avoiding a situation where some of the nodes
are heavily loaded while other nodes are idle or lightly loaded.

2
INTRODUCTION
Distributed systems offer the potential for sharing and aggregation of different resources such
as computers, storage systems and other specialized devices. These resources are distributed
and possibly owned by different agents or organization. The users of a distributed system
have different goals, objectives and strategies and their behavior is difficult to characterize. In
such systems the management of resources and applications is a very complex task. A
distributed system can be viewed as a collection of computing and communication resources
shared by active users.
When the demand for computing power increases, the load balancing problem becomes
important. The purpose of load balancing is to improve the performance of a distributed
system through an appropriate distribution of the application load. A general formulation of
this problem is as follows: given a large number of jobs, find the allocation of jobs to
computers optimizing a given objective function (e.g. total execution time).
Processing speed of a system is always highly intended. From the beginning of
the development of computer it is always focused on the system performance that is
how to improve the speed or performance of an existing system and thus we reached to
the era of supercomputer. Specially the business organizations, defense sectors and science
groups need the high performance systems for constant change of their day to day need. So
from serial computer to supercomputer, parallel computing and parallel distributed
computing have been developed. Massively parallel computers (MPC) are available in the
market today. In MPC a group of processors are linked to the memory modules through
the network like mesh, hypercube or torus. Super computers are very expensive so a new
alternative concept has emerged (although existed) that is called parallel distributed
computing in which thousands of processors can be connected either by wide area network or
across a large number of systems which consists of cheap and easily available autonomous
systems like workstations or PCs. So it is becoming extremely popular for large computing
purpose such as scientific calculations as compared to MPC. Recently distributed systems
with several hundred powerful processors have been developed. Distributed computing
system provides high performance environment that are able to provide huge processing

3
power. Multicomputer system can be efficiently used by efficient task partitioning and
balancing the tasks (loads) among the nodes properly .
Distributed network is mainly heterogeneous in nature in the sense that the processing
nodes, network topology, communication medium, operating system etc. may be different in
different network which are widely distributed over the globe. Presently several hundred
computers are connected to build the distributed computing system. In order to get the
maximum efficiency of a system the overall work load has to be distributed among the nodes
over the network. So the issue of load balancing became popular due to the existence of
distributed memory multiprocessor computing systems.
The distribution of loads to the processing elements is simply called the load
balancing problem. In a system with multiple nodes there is a very high chance that some
nodes will be idle while the other will be over loaded. The goal of the load balancing
algorithms is to maintain the load to each processing element such that all the processing
elements become neither overloaded nor idle that means each processing element ideally has
equal load at any moment of time during execution to obtain the maximum performance
(minimum execution time) of the system. So the proper design of a load balancing algorithm
may significantly improve the performance of the system.
In the network there will be some fast computing nodes and slow computing nodes. If we do
not account the processing speed and communication speed (bandwidth), the performance
of the overall system will be restricted by the slowest running node in the network. Thus load
balancing strategies balance the loads across the nodes by preventing the nodes to be idle and
the other nodes to be overwhelmed. Furthermore, load balancing strategies removes
the idleness of any node at run time.

4
LOAD BALANCING
Load balancing is the way of distributing load units (jobs or tasks) across a set of processors
which are connected to a network which may be distributed across the globe. The excess load
or remaining unexecuted load from a processor is migrated to other processors which have
load below the threshold load [9]. Threshold load is such an amount of load to a processor
that any load may come further to that processor. In a system with multiple nodes there is a
very high chance that some nodes will be idle while the other will be over loaded. So the
processors in a system can be identified according to their present load as heavily loaded
processors (enough jobs are waiting for execution), lightly loaded processors(less jobs
are waiting) and idle processors (have no job to execute). By load balancing strategy it is
possible to make every processor equally busy and to finish the works approximately at the
same time.
A load balancing operation consists of three rules. These are location rule, distribution rule
and selection rule The selection rule works either in preemptive or in non- preemptive
fashion. The newly generated process is always picked up by the non-preemptive rule while
the running process may be picked up by the preemptive rule. Preemptive transfer is costly
than non-preemptive transfer which is more preferable. However preemptive transfer is more
excellent than non-preemptive transfer in some instances.

5
Benefits of Load balancing
a) Load balancing improves the performance of each node and hence the overall system
performance.
b) Load balancing reduces the job idle time
c) Small jobs do not suffer from long starvation
d) Maximum utilization of resources
e) Response time becomes shorter
f) Higher throughput
g) Higher reliability
h) Low cost but high gain
i) Extensibility and incremental growth

6
Static Load Balancing
In static algorithm the processes are assigned to the processors at the compile time according
to the performance of the nodes. Once the processes are assigned, no change or reassignment
is possible at the run time. Number of jobs in each node is fixed in static load
balancing algorithm. Static algorithms do not collect any information about the nodes.
The assignment of jobs is done to the processing nodes on the basis of the following
factors: incoming time, extent of resource needed, mean execution time and
inter-process communications.
Since these factors should be measured before the assignment, this is why static load
balance is also called probabilistic algorithm. As there is no migration of job at the runtime
no overhead occurs or a little over head may occur.
Since load is balanced prior to the execution, several fundamental flaws with static load
balancing even if a mathematical solution exist: Very difficult to estimate accurately the
execution times of various parts of a program without actually executing the parts.
Communication delays that vary under different circumstances Some problems have an
indeterminate number of steps to reach their solution.
In static load balancing it is observed that as the number of tasks is more than the
processors, better will be the load balancing.
Fig shows a schematic diagram of static load balancing where local tasks arrive at the
assignment queue. A job either be transferred to a remote node or can be assigned to
threshold queue from the assignment queue. A job from remote node similarly be assigned to
threshold queue. Once a job is assigned to a threshold queue, it can not be migrated to any

7
node. A job arriving at any node either processed by that node or transferred to another node
for remote processing through the communication network. The static load balancing
algorithms can be divided into two sub classes: optimal static load balancing and sub optimal
static load balancing.
Model of Processing Node

8
Dynamic Load Balancing
During the static load balancing too much information about the system and jobs must
be known before the execution. These information may not be available in advance. A
thorough study on the system state and the jobs quite tedious approach in advance. So,
dynamic load balancing algorithm came into existence. The assignment of jobs is done at
the runtime. In DLB jobs are reassigned at the runtime depending upon the situation that is
the load will be transferred from heavily loaded nodes to the lightly loaded nodes. In this
case communication over heads occur and becomes more when number of processors
increase.
In dynamic load balancing no decision is taken until the process gets execution. This
strategy collects the information about the system state and about the job information.
As more information is collected by an algorithm in a short time, potentially the algorithm
can make better decision [10]. Dynamic load balancing is mostly considered in
heterogeneous system because it consists of nodes with different speeds, different
communication link speeds, different memory sizes, and variable external loads due to the
multiple. The numbers of load balancing strategies have been developed and classified so far
for getting the high performance of a system.
Fig shows a simple dynamic load balancing for transferring jobs from heavily loaded to the
lightly loaded nodes.

9
COMPARISON BETWEEN SLB and DLB ALGORITHM
Some qualitative parameters for comparative study have been listed below.
1. Nature
Whether the applied algorithm is static or dynamic is determined by this factor.
2. Overhead Involved
In static load balancing algorithm redistribution of tasks are not possible and there is
no overhead involved at runtime. But a little overhead may occur due to the inter process
communications. In case of dynamic load balancing algorithm redistribution of tasks are
done at the run time so considerable over heads may involve. Hence it clear that SLB
involves a less amount of overheads as compared to DLB.
3. Utilization of Resource
Though the response time is minimum in case of SLB, it has poor resource utilization
capability because it is impractical to get all the submitted jobs to the corresponding
processors will completed at the same time that means there is a great chance that some
would be idle after completing their assigned jobs and some will remain busy due to the
absence of reassignment policy. In case of dynamic algorithm since there is reassignment
policy exist at run time, it is possible to complete all the jobs approximately at the same time.
So, better resource utilization occurs in DLB.
4. Thrashing or Process Dumping
A processor is called in thrashing if it is spending more time in migration of jobs than
executing any useful work [1]. As the degree of migration is less, processor thrashing will be
less. So SLB is out of thrashing but DLB incurs considerable thrashing due to the process
migration during run time.
5. State Woggling
It corresponds to the frequent change of the status by the processors between low and high. It
is a performance degrading factor.

10
6. Predictability
Predictability corresponds to the fact that whether it is possible to predict about the behavior
of an algorithm. The behavior of the SLB algorithm is predictable as everything is known
before compilation. DLB algorithm’s behavior is unpredictable, as everything is done at run
time.
7. Adaptability
Adaptability determines whether an algorithm will adjust by itself with the change of
the system state. SLB has no ability to adapt with changing environment. But DLB has that
ability.
8. Reliability
Reliability of a system is concerned with if a node fails still the system will work without any
error. SLB is not so reliable as there is no ability to adapt with the changing of a system’s
state. But DLB has adaptation power, so DLB is more reliable.
9. Response Time
Response time measures how much time is taken by a system applying a particular
load balancing algorithm to respond for a job. SLB algorithm has shorter response time
because processors fully involved in processing due to the absence of job transferring.
But DLB algorithm has larger response time because processors can not fully involved in
processing due to the presence of job transferring policy.
10. Stability
SLB is more stable as every thing is known before compilation and work load transfer is
done. But DLB is not so stable as SLB because it involves both the compile time assignment
of jobs and distribution of work load as needed.
11. Complexity Involved
SLB algorithms are easy to construct while DLB algorithms are not so easy to develop
because nothing is known in advance. Although the dynamic load balancing is complex
phenomenon, the benefits from it is much more than its complexity [10].

11
COMPARISON of SOME DYNAMIC LOAD BALANCING
ALGORITHMS
Now some dynamic load balancing algorithms are studied below.
1. Nearest Neighbor Algorithm
With nearest neighbor algorithm each processor considers only its immediate neighbor
processors to perform load balancing operations. A processor takes the balancing decision
depending on the load it has and the load information to its immediate neighbors. By
exchanging the load successively to the neighboring nodes the system attains a global
balanced load state. The nearest neighbor algorithm is mainly divided into two categories
which are diffusion method and dimension exchange method.
2. Random (RAND) Algorithm
As soon as a workload (greater than threshold load) is generated in a processor, it is migrated
to a randomly selected neighbor. It does not check state information of anode. This
algorithm neither maintains any local load information nor sends any load information to
other processors . Furthermore, it is simple to design and easy to implement. But it causes
considerable communication overheads due to the random selection of lightly loaded
processor to the nearest neighbors.
3. Adaptive Contracting with Neighbor (ACWN)
As soon as the workload is newly generated, it is migrated to the least loaded nearest
neighbor processor. The load accepting processor keeps the load in its local heap. If the load
in its heap is less than to its threshold load then no problem otherwise it sends the load to
the neighbor processor which has load below the threshold load. So, ACWN does require
maintaining the local load information and also the load information of the neighbors for
exchanging the load periodically. Hence, RAND is different form the ACWN in a respect that
ACWN always finds the target node which is least loaded in neighbors.

12
4. Prioritized Random (PRAND) Algorithm
In both RAND and ACWN the work load is supposed to be uniform in the sense of their
computational requisites. Modification is done on RAND and ACWN for the non-
uniform workload to get prioritized RAND (PRAND) and prioritized ACWN (PACWN)
respectively. In these algorithms the work loads are assigned index numbers on the basis of
the weight of their heaps. PRAND is similar to RAND except that it selects the second
largest weighted load from the heap and transfers it to a randomly selected neighbor. On the
other hand, PACWN selects the second largest weighted workload and transfer it to the least
loaded neighbor.
4.1 Averaging Dimension Exchange (ADE)
ADE algorithm follows the concept of local averaging load distribution operation and
balances the workload with one of its neighbors.
4.2 Averaging Diffusion (ADF) Algorithm
It averages the local load distribution as ADE but exchanges the load with all its neighbors.
5. CYCLIC Algorithm
This is the outcome of RAND algorithm after slight modification. The workload is assigned
to a remote system in a cyclic fashion. This algorithm remembers always the last system to
which a process was sent.
6. PROBABILISTIC
Each node keeps a load vector including the load of a subset of nodes. The first half of the
load vector holding also the local load is sent periodically to a randomly selected
node. Thus information is revised in this way and the information may be spread in the
network without broadcasting. However, the quality of this algorithm is not ideal, its
extensibility is poor and insertion is delayed.
7. THRESHOLD and LEAST
They both use a partial knowledge obtained by message exchanges. A node is
randomly selected for accepting a migrated load in THRESHOLD. If the load is below

13
threshold load, the load accepted by there. Otherwise, polling is repeated with another node
for finding appropriate node for transferring the load. After a maximum number of attempts if
no proper recipient has been reported, the process is executed locally. LEAST is an instant of
THRESHOLD and after polling least loaded machine is chosen for receiving the migrated
load. THRESHOLD and LEAST have good performance and are of simple in nature.
Furthermore, up-to-date load values are used by these algorithms.
8. RECEPTION
In this algorithm, nodes having below the threshold load find the overloaded node by random
polling for migrating load from overloaded node.
9. Centralized Information and Centralized Decision
In this class of algorithms the information about the system is stored in a single node and the
decision is also taken by that single node. CENTRAL is a subclass of this algorithm. When a
heavily loaded node wants to migrate a job, it requests a server for a lightly loaded node.
Every node in the system informs the server machine whether a lightly loaded node is
available or not. CENTRAL afford very efficient performance results. But this algorithm
suffers from a very serious problem that if the server is crashed, no facility will be provided
by this algorithm.
10. Centralized Information and Distributed Decision
In GLOBAL, collection of information is centralized while decision making is distributed.
The load situation on the nodes is broadcasted by the server. Through this information an
overloaded processor finds the lightly loaded node from its load vector without going through
the server. This algorithm is very efficient due to the less inclusion of message information
and it is robust in nature because the system remains alive even when the server is in
recovery state. GLOBAL algorithm gathers large information but information is not up-to-
date. As a result greater overheads occur in the system.
11. Distributed information and Distributed Decision
Each node in OFFER broadcasts its load situation periodically and each node keeps a global
load vector. Performance of this algorithm is poor.

14
12.RADIO
In RADIO, both the information and decision are distributed and there is no
broadcasting without compulsion. In this algorithm, a distributed list consisting of lightly
loaded nodes in which each machine is aware of its successor and predecessor. Furthermore,
each node is aware of the head of the available list that is called manager. The migration
of a process from a heavily loaded node to the lightly loaded node is done directly or
indirectly through the manager. Broadcasting occurs when manager crashes or a node joins
the available list.
13. The Shortest Expected Delay (SED) Strategy
These strategy efforts to minimize the expected delay of each job completion so the
destination node will be selected in such a way that the delay becomes minimal. This is a
greedy approach in which each job does according to its best interest and joins the queue
which can minimize the expected delay of completion. The average delay of a given batch of
jobs with no further successive arrival is minimized by this approach. SED does not minimize
the average delay for an ongoing arrival process. To find out the destination node the source
node has to get state information from other nodes for location policy.
14. The Never Queue (NQ) Strategy
NQ policy is a separable strategy in which the sending server estimates the cost of sending a
job to each final destination or a subset of final destinations and the job is placed on the
server with minimal cost . This algorithm always places a job to a fastest available server.
This algorithm minimizes the extra delay into successive arriving jobs so that the overall
delay will be minimized by NQ policy. Furthermore, a server does not transfer incoming job
to the servers until fastest server than it is available.
15. Greedy Throughput (GT) Strategy
This strategy is different from SED and NQ strategies. GT strategy deals with the throughput
of the system that is the number of jobs completed per unit time would be maximum before
the arrival of new job instead of maximizing only the throughput rate at the instant of
balancing. This is why it is called Greedy Throughput (GT) policy [11, 16].

15
COMPARISON of DLB
The comparison of above discussed dynamic load balancing algorithms has been shown :
S.N. Algorithms State information
check
Performance
1 RAND No Excellent
2 PRAND Partial Excellent
3 ACWN Yes Good
4 PACWN Yes Good
5 CYCLIC Partial Slightly better than
RAND6 PROBALISTIC Partial Good
7 THRESHOLD Partial Better
8 LEAST Partial Better
9 RECEPTION Partial Not so good
10 CENTRAL Yes Excellent
11 GLOBAL Yes Good
12 OFFER Yes Poor
13 RADIO Yes Good
14 SED Yes Good
15 NQ Yes Good

16
CONCLUSION
In this paper we studied the load balancing strategies lucidly in detail. Load
balancing in distributed systems is the most thrust area in research today as the demand of
heterogeneous computing due to the wide use of internet. More efficient load balancing
algorithm more is the performance of the computing system. We made a comparison
between SLB and DLB introducing some new parameters. We have enumerated the
facilities provided by load balancing algorithms. Finally, we studied some important
dynamic load balancing algorithms and made their comparison to focus their importance in
different situations. There exists no absolutely perfect balancing algorithm but one can use
depending one the need.
The comparative study not only provides an insight view of the load balancing algorithms,
but also offers practical guidelines to researchers in designing efficient load balancing
algorithms for distributed computing systems.

17
REFERENCES
[1] Md. Firoj Ali and Rafiqul Zaman Khan. “The Study On Load Balancing Strategies In
Distributed Computing System”. International SJournal of Computer Science & Engineering
Survey (IJCSES) Vol.3, No.2, April 2012
[1] Ahmad I., Ghafoor A. and Mehrotra K. “Performance Prediction of Distributed
Load Balancing on Multicomputer Systems”. ACM, 830-839, 1991.
[2] Antonis K., Garofalakis J., Mourtos I. and spirakis P. “A Hierarchical Adaptive
Distributed Algorithm for Load Balancing”. Journal of Parallel and Distributed Computing,
Elsevier Inc.2003.
[4] E. Altman, T. Basar, T. Jimenez, and N. Shimkin. Routing in two parallel links: Game-
theoretic distributed algorithms. J. Parallel and Distributed Computing, 61(9):1367–1381,
September 2001.

Load balancing in Distributed Systems

More Related Content

What's hot (20)

Similar to Load balancing in Distributed Systems (20)

More from Richa Singh (7)

Recently uploaded (20)

Load balancing in Distributed Systems