Research Initiation: Dependability and Performance Models for Parallel Computers

Fast and accurate characterization of parallel computers is essential in finding their suitability for different applications. The thrust of the proposed research is to analyze and evaluate the dependability and performance issues of two promising classes of parallel computers. These are multistage interconnection network (MIN) systems and hypercube systems. The research is aimed at developing complete system models considering the degradation of both computing elements and communication network. Various dependability and performance measures, applicable in the domain of parallel processing, are used to characterize these architectures. Analytical/simulation/hybrid models are being developed to compute the dependability and performance-related dependability parameters for both repairable and non-repairable systems. The analytical models are based on approximate techniques and Markovian approach. The Markov model will be implemented as a program package to study the transient and steady state behavior of these systems. While the structure of the package will be somewhat similar to those of CARE III/HARP, it will have additional features such as automatic state generation and performance prediction. A major contribution of this research will be the availability of a set of tools that can be used for the evaluation of some of the existing and evolving parallel architectures.