The performance evaluation of workstation clusters

The requirement for High Performance Computing (HPC) have increased dramatically over the years. Parallelism is the only key technology today which can deliver the required computing performance for very large scale scientific and commercial applications, although its implementation in practice has proved to be a far more difficult task than originally envisaged.Advances in microprocessor and networking technologies in conjunction with the development and standardisation of the message-passing model and widespread availability of distributed software have enabled workstation clusters to have the potential for HPC at an attractive price-performance ratio. This combination of technologies provides several advantages for clusters but at the same time their evolution and performance is determined and limited by technologies designed for other systems. As a result clusters often fail to deliver at the application level their underlying potential performance.This thesis investigates the key components of commodity workstation clusters and evaluates the performance of these systems as an integrated HPC platform. It demonstrates the need for a new performance evaluation tool, and proposes the Specific Cluster Operation and Performance Evaluation (SCOPE) benchmark set which has been especially designed to evaluate the performance behaviour of cluster characteristics and promote the workstation cluster concept by assisting commodity workstation cluster designers to understand and analyse the performance behaviour of these systems.An initial implementation of the SCOPE benchmark suite has been developed and run on a wide variety of workstation clusters and MPP platforms. Results from the SCOPE tests have demonstrated the potential to identify and classify the performance evaluation of workstation clusters. Moreover the SCOPE evaluation tool methodology can be expanded to provide support for the development of parallel applications and algorithms tailored to a specific parallel platform.