A simulation study of dynamic load balancing for network-based parallel processing

A simulation study of dynamic load balancing for parallel processing on network of workstations (NOW) is presented in this paper. A simulation model is constructed. It includes a representative CPU scheduling policy, and also considers the message exchange, task transfer and migration costs explicitly. A global dynamic load balancing algorithm is simulated. Both task transfer and task migration are considered. The performance of the algorithm under both homogeneous and heterogeneous environments is analyzed. In addition, the interaction of parallel and sequential workloads on an NOW is also examined. Our results show that dynamic load balancing can achieve better performance improvement for heterogeneous systems than for homogeneous systems; it is especially effective in a system where both parallel and sequential tasks concurrently exist; and the use of task migration generally does not further improve performance