Performance assessment of network protocols and parallel programming tools for distributed computing systems

Abstract

Current advances in processor technology, networking technology, and software tools have made high-performance distributed computing over Local Area Networks (LANs) a promising alternative to traditional parallel and vector supercomputers. In this paper, we analyze the performance of two widely used distributed computing software tools, Express and PVM, on clusters of SUN workstations connected by an Ethernet and clusters of HP workstations connected by a Fiber Distributed Data Interface (FDDI) ring. Our goal is to assess the potential and pin-point the weaknesses of the various factors that affect the performance of these distributed computing environments. The major factors that are investigated in this paper are: networks, workstations, and distributed software tools. Extensive timing experiments, including one-to-one communications, exchange operations, and broadcast operations, have been performed in order to quantitatively determine the impact of these factors. Moreover, analytic models have been developed to analyze the behavior of the network protocols employed by the LAN-based platforms as well as to estimate the communication overhead for the distributed computing software tools.