Design and Implementation of Message Passing Interface for Clusters of Workstations on Local Area Networks

Abstract

Message passing is a paradigm widely used for writing parallel programs on clusters of workstations that have become an attractive and efficient alternative for parallel computing. Message Passing Interface (MPI) is a standard proposed for writing portable message-passing parallel programs. In this paper, we present a design and implementation of MPI that is optimized for LAN clusters of workststions. We develop a daemon-based implementation that consists of two main componenets: a MPI library that provide the functionality of MPI, and a multicast daemon that handles all message passing among MPI processes and provides multicast operations. We propose a user-level reliable multicast protocal, and implement it with sliding window technoques and hardware broadcast to improve performance. Performance measurement shows that our implementation outperforms two of the major implementations, MPICH and LAM, that are available from publib domain. Furthermore, our implementation allows uncoordinated checkpointing that is simple and efficient.