A hybrid load sharing algorithm for distributed memory systems

Load sharing strategies are aimed at reducing the average response time of tasks on a distributed system. In this paper, we study the problem of load sharing on a distributed memory system, and present a low overhead, hybrid algorithm which performs well both with heavy and light system load conditions. The solution we propose uses a fully distributed strategy when the system is heavy loaded, and a centralized one when part of the system become idle. The transition from one strategy to the other is automatically performed, so that there are no bottlenecks in the system. Both strategies are very lightweight and have very little overhead during both heavy and light load conditions. Also, they do not present instabilities under all load conditions. The performance of the hybrid algorithm has been simulated, and simulation results are presented under a variety of operating conditions