Giraphx: Parallel Yet Serializable Large-Scale Graph Processing

Bulk Synchronous Parallelism (BSP) provides a good model for parallel processing of many large-scale graph applications, however it is unsuitable/inefficient for graph applications that require coordination, such as graph-coloring, subcoloring, and clustering. To address this problem, we present an efficient modification to the BSP model to implement serializability (sequential consistency) without reducing the highly-parallel nature of BSP. Our modification bypasses the message queues in BSP and reads directly from the worker's memory for the internal vertex executions. To ensure serializability, coordination is performed--implemented via dining philosophers or token ring-- only for border vertices partitioned across workers. We implement our modifications to BSP on Giraph, an open-source clone of Google's Pregel. We show through a graph-coloring application that our modified framework, Giraphx, provides much better performance than implementing the application using dining-philosophers over Giraph. In fact, Giraphx outperforms Giraph even for embarrassingly parallel applications that do not require coordination, e.g., PageRank.