EPGraph: An Efficient Graph Computing Model in Persistent Memory System

Abstract

With the wide range of requirements to analyse large-scale graphs in many real-world applications (e.g., relationship analysis, fraud detection and product recommendation), graph computing recently receives intensive interests. However, the massive volume and the power-law distribution of graphs are objective obstacles to efficient graph computing. Fortunately, Intel Optane DC Persistent Memory (PMEM) has emerged as a new solution, which is expected to play a crucial role in large-scale graph processing. But compared with main memory, PMEM shows much lower bandwidth and higher access latency. Therefore, it becomes paramount to fully exploit the advantages of PMEM in persistent memory systems. In this paper, we propose EPGraph, a novel efficient graph computing model designed by PMEM. To a considerable extent, it improves the spatial locality and the temporal locality of graph computing at the same time. The main contribution of our work lies in three aspects. Firstly, we design a degree-based data layering strategy to reduce the impact of power-law distribution. The hierarchical strategy makes full use of DRAM and PMEM simultaneously. Secondly, we propose a dynamic migration mechanism during the iterative execution of graph computing. The dynamic mechanism equitably schedules the subgraphs which are used in the next iteration. Thirdly, we evaluate the effectiveness of EPGraph on five public graph data sets. Extensive evaluation results show that EPGraph outperforms state-of-the-art graph computing systems by 22.67%-35.03%.