Speculative Computation: Overcoming Communication Delays

Abstract

Communication latencies and delays are a major source of performance degradation in parallel computing systems. It is important to "mask" these communication delays by overlapping them with useful computation in order to obtain good parallel performance. This article proposes speculative computation as a technique to mask communication latencies in synchronous iterative algorithms. Processors speculate the contents of messages that are not yet received and perform computation based on the speculated values. When the messages are received, they are compared with the speculated values and, if the error is unacceptable, the resulting computation is corrected or recomputed. If the error is small, the speculated value is accepted and the processor has masked the communication delay. The technique, applied to N-body simulations yielded a performance improvement of up to 34%.