Energy Aware Load Prediction for Cloud Data Centers

Abstract

Amazon recently estimated that the cost of energy for its datacenters reached 42% of the total cost of operation. Our previous research proposed an algorithm to predict how much cloud workload is expected at a specific time. This allows physical servers determined not to be needed to be placed in a low-power sleep state to save energy. If more system capacity is required, servers in a sleep state are transitioned back to an active state. In this paper, we extend our prior research by presenting both a stochastic model for state change as well as a new approach to determining the sampling frequency for performing the prediction of the expected capacity. The first result we show is that this allows the optimal prediction time horizon to be chosen. We next present a dynamic prediction quantization method to determine the optimal number of prediction calculation intervals. Both of these new algorithms allow us to predict future load within required Service Level Agreements while minimizing the number of prediction calculations. This effectively optimizes our ability to predict while minimizing the detrimental effect of additional calculations on our cloud resources. Finally, we test this model by simulating the stochastic time horizon and dynamic quantization algorithms and compare the results with three competing methods. We show that our model provides up to a 20% reduction in the number of calculations required while maintaining the given Service Level Agreement.