On the performance of decentralized software

Abstract

Distribution of computing to achieve goals such as enhanced reliability depend on the use of decentralized software. Decentralization typically replaces a sequential process by a system of small, concurrent processes that interact frequently. The implementation of processes and their interactions represents a cost incurred as a result of decentralization. Performance measurements are reported in this paper for decentralized software written in a programming language for distributed computer systems. These performance measurements confirm that low-cost implementations of concurrency are possible, but indicate that decentralized software makes heavy use of run-time functions managing concurrency. An initial model comparing the performance of a specific decentralized software structure to its centralized counterpart indicates that these implementation costs are generally offset by the performance improvements that are due to the parallelism inherent in the decentralized structure. The research facilities for continued study of decentralized software performance are described in the summary.