Partitioning and scheduling loops on NOWs

Abstract

This paper addresses the problem of partitioning and scheduling loops for a network of heterogeneous workstations. By isolating the effects of send and receive and quantifying the impact of network contention on the overall communication cost, a simple yet accurate cost model for predicting the communication overhead for a pair of workstations is presented. The processing capacities of all workstations in a network are modeled based on their CPU speeds and memory sizes. Based on these models, loop tiling is used extensively to partition and schedule loops across the workstations. By adjusting sizes, i.e. the granularities of tasks, the impact of the heterogeneity arising from program, processor and network is minimised. Experimental results on an Ethernet of seven DEC workstations demonstrate the effectiveness of our models and parallelisation strategies.