Compression-aware graph computation

Abstract

Many recent work has focused on graph algorithms via parallelization including PowerGraph [9] and Ligra [14]. The frameworks process large graphs in shared memory, requiring a terabyte of memory and expensive maintenance cost. Reducing graph size to fit in memory thus is crucial in cutting the cost of large-scale graph computation. Compression has been widely used to reduce graph size. However, it could meanwhile compromise graph computation efficiency caused by nontrivial decompression overhead before graph computation. In this paper, we propose a simple and yet efficient coding scheme. It not only leads to smaller size of compressed graphs; meanwhile we can perform graph computation directly on the compressed graphs with no or partial decompression, namely compression-aware computation, leading to faster running time. Our experiments validate that the coding scheme achieves 2.99X compression ratio, and three compression-aware graph algorithms achieve 7.02X, 2.88X and 2.34X faster running time than the graph algorithms on the graphs without compression.