Power Optimized Load Balancing on a Cluster

Abstract

Maintaining an optimal performance on a parallel or distributed computing system requires load balancing. There are various techniques for load bal- ancing on a cluster. These can include job scheduling libraries or process migration libraries. We argue that just a job scheduling system or process migration system itself will not be enough to achieve a successful uniform workload distribution. So this project has proposed and implemented a pro- cess migration system that can run independently in a cluster, along with a job scheduling library. This will result in a more e ective work load dis- tribution, than using just a job scheduling or process migration library. In order to control the work load, this system monitors CPU temperature as an indirect measurement of its utilization. Then it tries to balance the work load by keeping the CPU temperature under a de ned maximum limit. Do- ing that, it will maintain an uniform workload distribution on the cluster. Uniform workload distribution will optimize the electric power utilization, while working towards the 'Green Computing' approach.