A bulk-synchronous parallel library implementation for the BBN butterfly GP1000

One of the fundamental goals of parallel computing is to develop a framework that will support portable and efficient application programs. The Bulk-Synchronous Parallel (BSP) model was proposed to help achieve this goal. The BSP model is intended to be a "unifying model"-it addresses both software and hardware issues by allowing theoretical analysis to coexist with practical physical implementations. For several years the BSP model has been supported mainly by theoretical results. Recent experiments, however, have begun to demonstrate the practicality of the model for real architectures running real applications. The goal of this paper is to describe the methodology used to construct an efficient BSP library on the BBN Butterfly GP1000. Our results are relevant for BSP library implementations on shared-memory systems in general and for NUMA (nonuniform m-memory-access) machines in particular.