Performance based reliability optimization for computer networks and distributed systems

A new reliability measure which incorporates connectivity and performance measures has been developed for computer networks as an alternative to conventional reliability or performance based measures. The measure developed in this dissertation, called the Average Network Throughput per unit installed capacity gives an indication of the traffic that can be reliably transferred between any pair of nodes in a network with unreliable components (the links) per unit installed capacity. Two new measures relating to distributed systems are also developed. They are Average Distributed Program Throughput per unit installed capacity and Average Distributed System Throughput per unit installed capacity. An algorithm to evaluate these measures is also developed. Examples have been provided to illustrate the effectiveness of these measures over conventional reliability measures. Two applications of these measures have been developed using the Genetic Algorithm. The first application deals with reliability improvement of an already existing distributed system and the second application deals with combined file and capacity allocation on a distributed system to maximize its reliability. In both applications a Genetic Algorithm based approach is used to optimize the appropriate performance related reliability parameter. The effectiveness of the Genetic Algorithm based approach is validated by comparing the GA results with the solutions obtained using an exhaustive search of the problem state space.