Automatic performance measurement of task allocation in a distributed system

Abstract

Distributed systems provide a big potential for improvement in job execution time with the aid of load balancing. In recent years, several theoretical and practical approaches have been taken to investigate the effects of load balancing. A lack of simple and adequate tools for this purpose was the main motivation for this thesis. This thesis studies the effects of load balancing on a distributed system using an experimental load balancer. It concentrates on the practical design and implementation of this load balancer and its transparent integration into an existing distributed environment, consisting of five Sun workstations running the UNIX operating system. Experimental results with the implemented load balancer show that although the remote execution overhead for initial placement is significant compared to the total job execution time, it is still possible to improve the total response time. A comparison of different scheduling algorithms shows that it is possible to improve the response time of standard UNIX jobs by over 40% on a moderately loaded system.