Right time in right place: Taming workload balancing oscillations in internet data center cost management

Cloud Service Providers (CSPs) have been spending a significant portion of their overall operational costs towards their electricity bills, to power up their Internet Data Centers (IDCs). Geographical Load Balancing (GLB) has been shown as an effective energy cost management solution for IDCs. However, due to the constantly changing dynamics of electricity price and cloud workload, an instantly optimal GLB decision may become suboptimal in the following time slots. An interesting finding in this paper reveals that most of the existing GLB frameworks share a previously undiscovered problem - the workload balancing oscillation problem. To solve this problem, we investigate the gap between the optimal energy cost and the costs with static GLB algorithms. To fill this gap, we propose a novel workload reallocation scheme and design an online cost minimization algorithm to implement this scheme in an eco-friendly electricity market. A bandwidth penalty is carefully imposed to control the workload balancing oscillations. Through performance analysis, we prove that the proposed algorithm can minimize the total cost and at the same time meet the desired tradeoff among cost, oscillations and delay. Extensive evaluations with real-world traces show that the proposed algorithm can reduce peak and total energy costs by up to 79.8% and 61.7% respectively, compared to static GLB algorithms.