Predictability of bulk synchronous programs using MPI

Abstract

The BSP cost model provides a general framework to design efficient and portable data-parallel algorithms. Execution costs of BSP programs are predicted combining a limited number of program and machine dependent parameters. BSP programs can be written using several programming tools. In this work we explore the predictability of bulk synchronous programs implemented with the Message Passing Interface. Two classic computational geometry problems: the convex hull (CH) and the lower envelope (LE) are considered as cases of study. Efficient BSP algorithms have been implemented using MPI and executed on three different parallel architectures: a Fujitsu AP1000 (distributed memory), a CRAY T3E (distributed shared memory) and a cluster of PCs (Backus). The paper compares the degree of predictability on these architectures, analysing the main sources of error.