A new heuristic for the process-processor mapping problem

This paper describes a new heuristic for the n-process p-processor task-assignment problem. The algorithm is based on repeatedly merging pairs of processes in order to reduce the initial n-node task interaction graph to a p-node one and at the same time trying to minimise the overall program execution time. We present the mapping algorithm and analyse its performance by comparing the results with optimal assignments. It is shown that for communication intensive problems running on a homogeneous network of processors, the algorithm produces acceptable suboptimal solutions in polynomial time, with errors less than 10% and typically below 5%.