Optimizing cost for geo-distributed storage systems in online social networks

Globally distributed data centers provide an opportunity to deploy geo-distributed Online Social Networks (OSNs). For so big data generated by users, how to store them among those data centers is a key issue in the geo-distributed storage system. Today's popular OSN providers store users' data in each deployed data center, so as to guarantee access latency. However, the full replication manner brings relatively high storage cost and traffic cost, which extremely increases the economic expenditure of OSN providers. Data placement based on social graph partitioning is an efficient way to minimize cost, but it requires the information of entire social graph and cannot fully guarantee latency. Recently, accomplished by partitioning replication is proposed to optimize cost as well as guarantee latency, but it has two drawbacks: (1) the separated manners of optimization cannot efficiently reduce the cost; (2) the placement of master replicas and slave replicas influence each other, and eventually reduces the optimization effects. In this paper, we explore an integrated manner of optimizing partitioning and replication simultaneously without distinguishing replica's role. We propose a lightweight replica placement (LRP) scheme, which conducts optimizations in a distributed manner and is well adapted to dynamic scenarios. Evaluations with two datasets from Twitter and Facebook show that LRP significantly reduces the cost compared with state-of-the-art schemes.