Building cost efficient cloud data centers via geographical load balancing

Abstract

Geographical load balancing (GLB) is widely established by cloud providers, to exploit the differences in electricity price, local green energy generation, transmission delay and cost etc across geographically dispersed data centers (DCs). With GLB, a cloud provider can achieve reduction of electricity cost or/and bandwidth cost or/and delay cost. However, these objectives are not independent from the others. In this paper, we first build an offline optimization-based framework to explore the trade-off among these three objectives under an emerging practical scenario, where the DCs are charged by peak pricing and burstable billing for energy consumption and bandwidth usage, respectively. It turns out that minimizing a specified cost may result in remarkable increases of other costs. Then, we propose a short-term prediction based mechanism (SPM) for the cloud provider to periodically deduce the request distribution strategies. With real-world data traces, we show that SPM can always obtain a suitable trade-off among different costs.