Performance models for I/O bound SPMD applications on clusters of workstations

Clusters of workstations represent today a satisfactory alternative to MPPs and supercomputers in many areas of application. The rapidly reduction of the cost of high performance workstations/PCs makes this technology ever more available. Moreover, new concepts for the integration of individual workstations through Local Area Networks are emerging. High speed interconnection networks and optimized protocol system architectures are the most important objectives of current research in this field of study. In this contribution, we attempt to propose a simple but effective performance model of systems with distributed computational and I/O resources when executing parallel scientific applications characterized by communication bursts and by intensive I/O phases. By means of queueing network techniques, the analysis of the model leads to the definition of a speedup surface which captures the relative influence of processors and disks parallelism in the performance of applications that alternates computations and I/O operations in a cyclic fashion.