High Performance Message Passing Library for Maestro2 Cluster Network.
Keiichi Aoki, Koichi Wada, Masaaki Ono, Shinichi Yamagiwa · Parallel and distributed computing and networks · 2005
A cluster has become a popular platform for high performance computing. Currently, many PC clusters are organized with cost-effective conventional wide-area network technologies to connect each processing elements. Here, it is difficult to extract potential performance of the cluster, since there is the disparity between the wide-area network technologies and the performance characteristics required by cluster computing. We have developed a network to build high performance PC cluster, called Maestro2 cluster network. This paper describes about high performance communication library for Maestro2, named MMP. First, we describe the novel technique, design, and implementation of Maestro2. Then, we present a design and an implementation of MMP. We also show experimental results to evaluate the basic performance of MMP and compare it with raw performance of Maestro2. The experimental results show that MMP can extract maximum performance from Maestro2.