Dynamic load balancing for parallel program execution on a message-passing multicomputer

Dynamic load balancing solves the remapping problem in a multicomputer system at run time, where many processes need to be allocated evenly to multiple processor nodes. The mean is to migrate processes from busy to idle nodes in order to achieve higher resource utilization. The authors have implemented a distributed load balancer at an iPSC/2 hypercube computer system, which uses heuristic methods to balance the system load adaptively. These heuristic methods cooperate with a central supervisor at the host machine, and invoke the load balancing activities under decentralized control. Benchmark experiments have shown that the proposed dynamic load balancing methods can speed up the parallel execution of benchmark programs significantly.<<ETX>>