Conducting Graph Analytics In Graph Database Systems By Using GPU-based Accelerators

In the big data era,graph databases store highly correlated large-scale data becomes a research hot spot.However,the traditional distributed graph computing system consumes a lot of time and causes the computing results to lose effect.Because of the massive degree of parallelism and the high memory access bandwidth in Graphics Processing Unit(GPU),GPU-based graph database computing system can improve the analysis efficiency of data in the graph database.We propose a graph system HugeRock based on graph database HugeGraph and GPU parallel programming model GunRock,it solves the problem of analyzing graph data efficiency and realizes the storage and computing functions of large-scale graph data.First,we build a data storage system with Hbase as the storage backend.Secondly,we design a tool to finish the work of loading the graph dataset,especially realworld graphs.Thirdly,we propose a remote data management system to remotely manage the graph database,and convert graph data to Compressed Sparse Row(CSR)format data which is friendly to GPU.Finally,the iterative computing function of graph data is designed and implemented by using the programming model on GPU.To examine the performance of the proposed system,we use four large-scale,real-world social graphs as dataset and test several representative graph algorithms with these dataset.Experimental results indicate that our system can effectively complete the graph processing tasks and achieve better performance than GraphX,the graph analysis system on GPU compute speed is 5.5-6 times that of the GraphX graph analysis system on Sprak.