Computation and Data Partitioning on Scalable Shared Memory Multiprocessors

In this paper we identify the factors that affect the derivation of computation and data partitions on scalable shared memory multiprocessors (SSMMs). We show that these factors necessitate an SSMM-conscious approach. In addition to remote memory access, which is the sole factor on distributed memory multiprocessors, cache affinity, memory contention and false sharing are important factors that must be considered. Experimental evidence is presented to demonstrate the impact of these factors on performance using three applications on the KSR1 and the Hector multiprocessors. 1 Introduction Scalable shared memory multiprocessors (SSMMs) are becoming increasingly popular and a viable alternative to distributed memory multiprocessors (DMMs). The Stanford DASH [20], FLASH [14], the KSR1 [24], Toronto's Hector [26], NUMAchine [1], and the Cray T3D [23] are some SSMMs currently in use or under development. Processors in a SSMM share a single coherent address space. However, shared memory is p..