Analysis of computing resource allocation for replicated processes in P2P-MPI

摘要

In this paper, we explore the computing resources, or nodes, allocation for replicated processes in P2P-MPI. It uses the communication latency between the root node and other computing nodes to rank all nodes into a list, and selects the nodes which have shorter latency to root node to perform the computing task. The analyses and experiment shows that allocation scheme was far from optimized. If the latency and bandwidth between nodes which have communication during the computing was considered, a more effective allocation could be obtained. The analyses show that selecting the optimized allocation is an NP-hard problem with some difference from traditional parallel task allocation or scheduling. An alternative allocation scheme is proposed, and some opening problems are discussed.