Accelerating Depth-First Traversal by Graph Ordering

Abstract

Cache efficiency is an important factor in the performance of graph processing due to the irregular memory access patterns caused by the sparse nature of graphs. To increase the cache hit rate, prior studies proposed a variety of preprocessing approaches based on the reordering, which permutes the vertexes’ labels to improve the locality of graph structures. However, the locality enhancement of existing reordering approaches does not bring much performance benefit in depth-first traversal, which is widely adopted in a majority of graph processing applications. Furthermore, the state-of-the-art reordering approach suffers from an obvious overhead on preprocessing which will greatly limit the application of their approach. In this paper, we propose SeqDFS, a depth-first graph traversal method that optimizes the cache efficiency by adjusting the order of vertexes visited and can be further extended to dynamic scenarios. We conduct extensive experiments on 16 real-world datasets and 3 representative depth-first graph applications, of which the results show that our proposal achieves a significant speed-up on both directed and undirected graphs.