High performance computing over switch-based high-speed networks

Communication between processors has long been the bottleneck of distributed network computing. However, recent progress in switch-based high-speed networks including ATM, HIPPI and Fibre Channel may be changing this situation. To utilize the aggregate computational power of clusters of workstations, We have especially concentrated our effort on maximizing the achievable throughput and minimizing the delay for end-users. In order to achieve better performance. we have investigated the following issues: i) ways to improve the performance of I/O subsystems and ii) porting PVM over high-speed network APIs. The typical communications between processors are going through a stack of network protocols. (i) represents our effort to improve the lower layer communication performance. However, the application (or user) level of performance may not be improved too much even if the lower layer performance is improved. Therefore, to improve user level performance it is required to bypass most of the network protocols in the communication stack and, hence, the overheads generated by them by porting applications (like PVM) directly over lower layer network protocols. For (i) we discuss the details of how the I/O subsystems related to the network operations and several possible approaches for improving the maximum achievable bandwidth and reducing end-to-end communication latency. For (ii) we show some performance results of PVM over high-speed local area networks. Although the available bandwidth of high-speed network is much higher than the traditional LANs, the application-level performance, however, still lags behind the capabilities of the physical medium.