Dynamic load balancing on a distributed system

Abstract

We consider the problem of load balancing on loosely coupled multiprocessor systems. During run time, a task may create subtasks, which are dynamically distributed by the load balancer. Different load-balancing strategies (receiver-initiated, sender-initiated and mixture of both) are studied and evaluated on Transputers. We test three commonly used benchmark problems (fibonacci function, N-queen and 15-puzzle) to observe the effect of load balancing. Our experiments involve up to 18 Transputers, and we observe speed improvements from 12 to 16 times over a sequential program. The mixed strategy was the best in most cases. We also find that the longer a problem takes to solve using sequential implementation, the more likely it is to benefit from parallel execution. The load balancing algorithms presented here are applicable to any distributed systems where processor interconnection is modifiable.