An evaluation of parallel algorithms on current memory consistency models

Abstract

Memory consistency model is crucial to the performance of shared-memory multiprocessors, and in current architectures several different models are adopted. In this paper, using graph algorithms for illustrative purposes, we consider the impact of memory model on the implementation and performance of parallel algorithms on shared-memory multiprocessors. We show that the implementation of PRAM algorithms is largely "oblivious" of the underlying memory model, and has good performance on relaxed models. More importantly, we show that different memory models can favor drastically different algorithm designs.