Cost Optimization for Dynamic Replication and Migration of Data in Cloud Data Centers

Cloud Storage Providers (CSPs) offer geographically data stores providing several storage classes with different prices. An important problem facing by cloud users is how to exploit these storage classes to serve an application with a time-varying workload on its objects at minimum cost. This cost consists of residential cost (i.e., storage, Put and Get costs) and potential migration cost (i.e., network cost). To address this problem, we first propose the optimal offline algorithm that leverages dynamic and linear programming techniques with the assumption of available exact knowledge of workload on objects. Due to the high time complexity of this algorithm and its requirement for a priori knowledge, we propose two online algorithms that make a trade-off between residential and migration costs and dynamically select storage classes across CSPs. The first online algorithm is deterministic with no need of any knowledge of workload and incurs no more than $2\gamma -1$2γ-1 times of the minimum cost obtained by the optimal offline algorithm, where $\gamma$γ is the ratio of the residential cost in the most expensive data store to the cheapest one in either network or storage cost. The second online algorithm is randomized that leverages “Receding Horizon Control” (RHC) technique with the exploitation of available future workload information for $w$w time slots. This algorithm incurs at most $1+\frac{\gamma }{w}$1+γw times the optimal cost. The effectiveness of the proposed algorithms is demonstrated through simulations using a workload synthesized based on characteristics of the Facebook workload.