Architectural support for high-performance distributed computing

The emergence of high speed networks and the proliferation of high performance workstations have attracted a lot of interest in workstation-based distributed computing. Current trend in local area networks is toward higher communication bandwidth as we progress from Ethernet networks Thai operate at 10 Mbit/sec to higher speed networks that can operate in Gbit/sec range. Also, current workstations are capable of delivering tens and hundreds of Megaflops of computing power. By using a cluster of such high-performance workstations and the high-speed networks, a high-performance distributed computing environment could be built in cost-effective manner as an alternative of supercomputing platform. Howeverf in current local area networks, the bandwidths achievable at the application level are often an order of magnitude lower than that provided at the network medium [3, 7/It is therefore not sufficient to have even a Gigabit data link if user applications could only use a small portion of that bandwidth. In this paper, we present a software and hardware support to transform a local area network of workstations into a high-performance distributed computing environment, We present a Host Interface Processor (HIP) and a communication protocol (HCP) in order to improve the application-level transfer rates. We also analyze the performance of a distributed application when it runs on the computers of the HIP-based local networks and compare it with the performance of a single computer execution,.