Performance Prediction and Network Media Selection for PVM Clusters

Workstation clusters are becoming more popular as a parallel computing platform. Several programming libraries, including Parallel Virtual Machine (PVM), allow workstation clusters to be programmed as a single parallel computer. The performance of parallel programs is often affected by complex interactions between hardware and software. Performance prediction tools can help characterize the performance of parallel programs for different problem sizes, numbers of processors, and hardware configurations. This research investigates the feasability of developing a performance prediction system for PVM applications running on workstation clusters. We have developed the APACHE (Automated Pvm Application CHaracterization Environment) performance prediction system for PVMprograms. We review the system implementationand present some experimental results. The performance of workstation clusters in parallel processing has traditionally been limited by the communication network. New networking technologies such as Asynchronous Transfer Mode (ATM) promise to relieve the performance bottleneck of workstation clusters. However, not all parallel applications will be helped by these new technologies. One objective of the current research is to investigate the use of information obtained from a performance prediction system as a basis for evaluating different network media for use with a particular set of parallel applications. COMMITTEEAPPROVAL: Mark J. Clement, Committee Chairman J. Kelly Flanagan, Committee Member Tony Martinez, Committee Member Tony Martinez, Graduate Coordinator