Minimizing the load redistribution cost in cluster architectures

Abstract

In this work, we deal with the problem of minimizing the load redistribution cost in parallel implementations for cluster architectures. Due to the importance of the network latency in this kind of systems, the redistribution cost is primarily depending on the maximum number of messages sent or received by a processor. The load redistribution is a NP-hard problem similar to the multiple knapsack problems. Three heuristics are proposed to solve the problem in a global context, and a comparison is made to emphasize their characteristics. In a parallel application, it is important to decide whether it is efficient or not to carry out the workload redistribution. This decision is taken comparing the cost of the load imbalance and the communication overheads associated with the load balancing heuristic. Depending on these costs, a theoretic value of imbalance from which the redistribution is profitable is defined. Experimental results show the accuracy of our proposals.