Abstract
Cluster computers applying personal computers (PC) to their processing nodes have been widely used as parallel computational platforms due to the improvements of cost/performance ratio. In order to achieve higher performance of a cluster computer, it is important to be considered not only an approach to improve the network performance but also the one to overlap computation with communication. To confirm the validity of the latter approach, we are developing the Maestro2 cluster network which is composed of generalβpurpose processors and dedicated communication hardware tightly coupled with those processors. This paper describes a design and implementation of a communication library of a message passing system, called MMP, over the Maestro2 cluster network. On MMP, the processor on the Maestro cluster network controls the hardware and processes almost all communication steps such as for sending/receiving messages and a synchronization among processing units. According to an experimental evaluation, this paper shows that MMP achieves higher throughput than the conventional communication method that is processed by a host processor. Β© 2007 Wiley Periodicals, Inc. Electron Comm Jpn Pt 2, 90(11): 109β121, 2007; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/ecjb.20411