Reliability and performance models for reconfigurable computer systems
Ravi Varadarajan · 1987
A reconfigurable computer system can change its physical, functional, architectural and other characteristics to improve its reliability, performance or both. Since different notions of reconfiguration exist, we have attempted to provide a formal definition of reconfiguration and reconfigurable computers systems. One of the main issues in reconfigurable computers is the design of reconfiguration algorithms which choose at times of reconfiguration, system configurations optimal with respect to some reliability and performance attributes. We propose two types of optimization for reconfiguration algorithms, namely Commodity distribution models and Sample path constrained Markov decision models. The commodity distribution are useful for load balancing with resource migration in distributed systems. When the bottleneck cost of migration is minimized, the model reduces to a bottleneck transportation problem. We demonstrate its application to the two cases: File migration in distributed databases and host migration in mobile computer networks. We have also developed an efficient algorithm to solve a special case of the bottleneck transportation problem. The constrained Markov decision are useful for reconfigurable fault-tolerant computers in which reconfiguration involves reliability, performance tradeoff. We propose path constraint which requires the time-average cost to be below a specified value with probability one. We have developed a novel decomposition theory to find an optimal or near optimal policy for the sample path problem. This approach lends itself to parallel processing.