Improving Performance of Clusters using Load Balancing Algorithms

Abstract

A Cluster is a type of parallel or distributed processing system, which consists of a collection of interconnected stand alone computers working together as a single integrated computing resource generally having a single system image (SSI) that is the users generally view the clusters as a single system. A Cluster can be used in scientific applications that need supercomputing power or in various other areas like databases, multimedia, web services, etc. In addition the users can access any node of the cluster. Load balancing tries to balance the total cluster system load by transferring or starting processes on the more lightly loaded nodes in preference to heavily loaded nodes.Many load balancing algorithms have been proposed such as round-robin (RR) scheduling, weighted round-robin (WRR) scheduling, least-connection (LC) scheduling and WLC (Weighted Least Connection) scheduling. Comparative studies on their performance have thrown light on their merits and drawbacks with regards to various parameters. This thesis deals with the study of various load balancing algorithms and an attempt at producing an algorithm that can improve the cluster performance.
The proposed algorithm can be viewed a two stage process to balance the load in Heterogeneous Clusters though it works perfectly for homogeneous clusters as well. In the first stage the nodes are analyzed to produce their performance factor and in the second stage the load balancing is done. The algorithm is validated with simulations done in MATLAB and comparing the results with existing algorithms for load balancing in clusters