AN ANALYTICAL STUDY OF MULTIPROCESSOR SYSTEM RELIABILITY
Abstract
In the last two decades, intensive research activities have been conducted on the development of multiprocessor system to achieve high system reliability, availability, throughput, and fault tolerance capabilities. The reliability performance of multiprocessor architectures is studied in this thesis. Multiprocessor systems being studied are gracefully degradable. A formal probabilistic-graph is proposed to model multiprocessor systems. The Multi-partite graph is used to model multiprocessor architectures. Three categories of multiprocessor systems are considered according to their interconnection structures. (i.e. configurations with crossbar switch, time-shared busses, and multiport memories). The multiprocessing performance level is defined by the resources demanding threshold. Combinatorial enumeration techniques are used to find the reliability expressions. Closed formulations are obtained to evaluate the system reliability/availability of all three multiprocessor architectures at different performance levels. The analytical results are effectively applied to evaluate commercial systems including the C.mmp, B6700,Tandem/16, MIT/IL ACGN, Univac 1100/80, IBM 370/168, and DEC-K110. The reliability sensitivity of various multiprocessor systems are studied to determine the most critical components in a given system. Two optimization techniques were developed to maximize the system reliability under a given cost constraint. The system reliability is optimized by either choosing optimal component reliability or by adding optimal redundancy to the system. Performance measures on the availability and Mean Time Before Failure of multiprocessing systems are obtained with the reliability study. Relative reliability and availability performances of all three multiprocessor architectures and three physical crossbar implementations are compared with in-depth discussions.
Degree
Ph.D.
Subject Area
Electrical engineering
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.