Performance modeling of computation and communication tradeoffs in vertex-centric graph processing clusters

Abstract

Distributed vertex-centric graph processing systems have been recently proposed to perform different types of analytics on large graphs. These systems utilize the parallelism of shared nothing clusters. In this work we propose a novel model for the performance cost of such clusters.We also define novel metrics related to the workload balance and network communication cost of clusters processing massive real graph datasets. We empirically investigate the effects of different graph partitioning mechanisms and their tradeoff for two different categories of graph processing algorithms.