Performance-related Dependability Evaluation of Supercomputer Systems

Jean ARLAT, J.-C. LAPRIE

Abstract

The paper presents an example of performance-related dependability evaluation of a supercomputer structure corresponding to an MIMD multiprocessor system intended for high speed scientific computation. The approach presented addresses the problems of deriving a model that is tractable yet representative of the behavior of a complex system. This is achieved by means of an intensive validation study, and through the evaluation of measures of interest which account for the specific operating requirements characterizing the system under investigation: a) maintain very high throughput over a long period of time, and b) provide an efficient operational life-cycle.

Keywords: Supercomputers, Omega network, degradable performance, dependability evaluation, performance evaluation