EVALUATING DEPENDABILITY OF MESH-CO 'ECTED PROCESSORS

Abstract

For a network topology of connected processors, the tenn fault means incorrect response or failure of a processor or communication link. A system is fault-tolerant [8] if it remains working though there may be some faults. In this thesis, we consider mainly an increasingly-important parallel processors topology, which is the mesh topology. An extension of the mesh topology, the torus one, is considered afterwards. Dependability is a generic tenn used to address reliability, availability, maintainability, and the related issues [1], [15]. In this thesis, we address the reliability evaluation of the mesh-connected as well as the torus-clonnected processors. Reliability is a measure offimlt-tolerance of connected-processors in terms of how large the mesh is, how large the required submesh is, and how long the mesh is expected to run before failure. The same applies to the torus topolgy. Reliability can be estimated in both topologies as the probability of having at least one submesh (subtorus) working for a demanding task. This can happen in single instruction stream multiple data stream (SIMD short), and multiple instruction stream multiple data stream (MIMD).