Reducing the Communication Costs of Graph Analysis by Read-Only Replicas and Prioritized Execution

Graph mining is widely used in fields like social network analysis. The synchronous vertex-centric frameworks strike better balance between the performance and ease-of-use, so they are widely used in realistic. However, traditional architectures of this type, like the vertex-based push architecture and GAS, are encumbered by high communication costs. In this paper we proposed a new replica-based push architecture, which relies on the read-only buffers to reduce the communication costs. Based on this new architecture, we further incorporated the prioritized policy which limits the vertex updating orders. The prioritized policy further decreases the workloads with the costs of more iterations, and is suitable for large-scale graphs. The experiments demonstrated the effectiveness of our design, which decreases the running time by 58% ~ 85% compared with present architectures.