Social Network-Aware Disk Management

Abstract

Disk access patterns of social networking applications are different from those of traditional applications. However, today's disk layout techniques are not adapted to social networking workloads and thus suffer in performance. In this paper, we first present disk layout techniques that leverage community structure in the social graph to make placement decisions. Second, we build a layout manager called the Bondhu system that incorporates our techniques. We integrate Bondhu into the popular Neo4j graph database engine. Our trace driven experimental results show that the Bondhu system improves the median response time by as much as 48%. While taking the community structure into account yields clear benefits, our results indicate that models with more complexity beyond the social graph may yield low additional benefit.