ELite: Cost-effective approximation of exploration-based graph analysis
Vertex-centric block synchronous processing systems, exemplified by Pregel and Giraph, have received extensive attention for graph processing. These systems allow programmers to think only about operations that take place at one vertex and provide the underlying computation framework that involves multiple iterations (supersteps) with communication between neighboring vertices between supersteps. As graphs grow in size to billions of vertices and trillions of edges, processing them in this model face challenges: (1) The poor latency of supersteps dominated by the tasks performed on high degree vertices or densely connected components; and (2) The overwhelming network communication among vertices that can be proved of high redundancy. For many applications, approximate results are acceptable, and if these can be computed rapidly, they may be preferable. Many of the existing approximate solutions suffer from algorithm-specific designs that are not generic or lacking theoretical guarantees on the results' quality. In this paper we tackle this problem using a generic approach that can be incorporated into the graph processing platform. The approach we advocate involves communicating vertex states to a subset of the neighbors at each superstep; this is called selective edge lookup. We show how this approach can be incorporated into two primitive graph operators: BFS and DFS, which can be the basis of many graph analysis workloads. Extensive experiments over real-world and synthetic graphs validate the effectiveness and efficiency of the selective edge lookup approach.
Proceedings of the 3rd ACM SIGMOD Joint International Workshop on Graph Data Management Experiences and Systems and Network Data Analytics, GRADES-NDA 2020
Zhang, X., Özsu, M., & Chen, L. (2020). ELite: Cost-effective approximation of exploration-based graph analysis. Proceedings of the 3rd ACM SIGMOD Joint International Workshop on Graph Data Management Experiences and Systems and Network Data Analytics, GRADES-NDA 2020 https://doi.org/10.1145/3398682.3399164