A Common Messaging Layer for MPI and PVM over SCI

This paper describes the design of a common message passing layer for implementing both MPI and PVM over the SCI interconnect in a workstation or PC cluster. The design is focused at obtaining low latency. The message layer encapsulates all necessary knowledge of the underlying interconnect and operating system. Yet, we claim that it can be used to implement such diierent message passing libraries as MPI and PVM without sacriicing eeciency. Initial results obtained from using the message layer in SCI clusters are presented. 1 Background and Motivation For several years now, workstation or PC clusters have been in use as highly cost-eeective platforms for parallel computing. However, to become real alternatives to expensive dedicated parallel machines, clusters must be equipped with high-speed interconnects and a suuciently complete software basis supporting parallel programming and enabling application porting with relative ease. The Scalable Coherent Interface (SCI) is a promising interconnect technology for building such clusters, facilitating high-throughput and, most importantly, low-latency communication in a cluster environment 7]. While hardware (SCI adapter cards and switches) and basic device driver software for SCI-based clusters have been on the market for about three years, standard parallel programming APIs over SCI have not yet become available. The ESPRIT HPCN project SISCI (Standard Software Infrastructure for SCI-based Parallel Systems) aims at establishing this missing link, by implementing and evaluating the following formal and de facto standard software environments on SCI clusters: i) the Message Passing Interface communication library (MPI); ii) the Parallel Virtual Machine parallel programming system (PVM); and iii) a POSIX compliant, distributed thread package (Pthreads). These packages will be evaluated by a number of demanding parallel applications. More information on SISCI is available in 3] and from 18].