Workload Decomposition for Power Efficient Storage Systems

Power consumption and cooling costs of hosted storage services and shared storage infrastructure in data centers account for a growing percentage of the total cost of ownership. The bursty nature of storage workloads requires significant over provisioning of the capacity and power consumption to meet traditional response-time QoS guarantees. In this paper, we propose a graduated, distribution-based QoS model and a runtime scheduler for power efficient storage servers. Through the evaluation of the real storage workload traces, we show a new general tradeoff between the performance and power consumption.