Cost efficient datacenter selection for cloud services

Many cloud services nowadays are running on top of geographically distributed infrastructures for better reliability and performance. They need an effective way to direct the user requests to a suitable datacenter, in a cost efficient manner. Previous work focused mostly on the electricity cost of datacenters. The approaches favor datacenters at locations with cheaper electricity prices. In this paper, we augment the picture by considering another significant cost contributor: network bandwidth. We propose to utilize statistical multiplexing to strategically bundle demands at different locations. The anti-correlation between demands effectively smooths out the aggregated bandwidth usage, thereby saving the bandwidth cost calculated by burstable billing methods that charge the peak bandwidth usage. We present an optimization framework that models the realistic environment and practical constraints a cloud faces. We develop an efficient distributed algorithm based on dual decomposition and the subgradient method, and evaluate its effectiveness and practicality using real-world traffic traces and electricity costs.