Traffic-Aware Data Placement for Online Social Networks

Recently, Online social networks (OSNs) are growing at a phenomenal rate, and OSN service providers need to consider how to place users´ data to multiple servers inside a data center. Key-value stores use consistent hashing to efficiently solve the problem of data placement, and turn into a defacto standard. However, random placement manner of hashing cannot preserve social locality, which leads to high intra-data center traffic as well as unpredictable response time. To preserve social locality, many existing works model the data placement problem as a graph partitioning problem. There are two flaws in these works: (1) The social graph or interaction graph is constructed with ordinary pairwise graph that cannot fully reflect multi-participant interactions occurred in OSNs. (2) The underlying network topologies of data center are not considered in previous works. In this paper, we explore traffic-aware data placement for data storage in OSNs while trying to maximally preserve both social locality and distance locality. We formulate the problem as two sub-problems - hypergraph partitioning and partition-to-server mapping, and propose a traffic-aware data placement (TDP) scheme. Through extensive experiments with a large scale real world Facebook trace, we evaluate that TDP significantly reduces intra-data center traffic without deteriorating load balance among servers.