An Empirical Study of Message-Passing Overhead
Michael Lee Scott, Alan L. Cox · UR Research (University of Rochester) · 1986
Conventional wiodom holds that me.sage-passing is orders of magnitude more expensive than shared memory for communication between parallel proce..... Differences in the speed of underlying hardware mechaniems fail to account for a substantial portion of the performance gap. The remainder is ganerally attributed to the "inevitable coat " of higher-level semantic., but a dseper understanding of the factors thet contribute to message-pa88ing overhead has not been forthcoming. 1n this paper we provide a detailed performance analysis of one message-p88sing ayBtem: the implementation for the BBN Butterfly Parallel Proce880r of the LYNX distributed programming language. The case study includes a description of the implementation, an explanation of optimizations employed to improve ita performance, and a detailed breakdown of remaining cost.. The data provide a direct measure of the expense of individual feature. in LYNX. They also provide insight into the likely cost. of other message-p888ing syBtems, both present and future. Le.sons gained from our experience should be of use to other researchers in performing similar studies. 1.