LESSONS LEARNED FROM IMPLEMENTING BSP

We focus on two criticisms of bulk synchronous parallelism (BSP): that delaying communication until specific points in a program causes poor performance, and that frequent barrier synchronisations are too expen sive for high-performance parallel computing. We show that these criti cisms are misguided, not just about BSP but about parallel programming in general, because they are based on misconceptions about the origin s of poor performance. The main implication for parallel programming i s that higher levels of abstraction do not only make software construc tion easier - they also make high-performance implementation easier. C opyright (C) 1998 Elsevier Science B.V.