Research Based on Communication Affects Performance of Cluster Parallel Computing

This paper introduces LogGP model and the algorithmic communication model, points out how communication parameters affect parallel performance. Considering parallel program communication characteristic, Presenting four frequently used methods to improve clusters parallel computing performance, including using faster network, adopting user-level network interface protocols, using SMPs as computing nodes, dynamic prefetching data or migrating data home, and merging small messages. The performance of each method is strictly tested and the characteristic of each one is anatomized. Using faster network to reduce network latency is common and effective to all kinds of parallel applications. Those that send a great number of small messages should adopt user-level network interface protocols additionally to reduce spending. Merging small messages is most efficient for particular applications. Be careful to rep fetch data dynamically or migrate data home, for the optimization is suitable only for some applications. Don't start too many processes in one node w hen using SMP s to run the program.