Neural networks for the design of distributed, fault-tolerant, computing environments

Abstract

Binary optimization models for the design of distributed, fault-tolerant computing systems are considered, with a focus on the task allocation and file assignment modeling schema proposed by J. Bannister and K. Trivedi (Proc. Second Symp. on Reliability in Distributed Software and Database Systems, 1982). It is shown that R. Graham's (1969) partitioning algorithm, S, when applied to this schema can, in the case of finite resources, yield allocations that are arbitrarily poor with respect to the optimum allocation. This contrasts sharply with the case of ample resources, where S provides allocations that are provably close to the optimum. Two alternative allocation algorithms are suggested. Both are seen to deliver allocations preferable to those provided by S, but at some additional computational expense.<>