PCAP: Performance-Aware Power Capping for the Disk Drive in the Cloud

Power efficiency is pressing in today's cloud systems. Datacenter architects are responding with various strategies, including capping the power available to computing systems. Throttling bandwidth has been proposed to cap the power usage of the disk drive. This work revisits throttling and addresses its shortcomings. We show that, contrary to the common belief, the disk's power usage does not always increase as the disk's throughput increases. Furthermore, throttling unnecessarily sacrifices I/O response times by idling the disk. We propose a technique that resizes the queues of the disk to cap its power. Resizing queues not only imposes no delays on servicing requests, but also enables performance differentiation. We present the design and implementation of PCAP, an agile performance-aware power capping system for the disk drive. PCAP dynamically resizes the disk's queues to cap power. It operates in two performance-aware modes, throughput and tail-latency, making it viable for cloud systems with service-level differentiation. We evaluate PCAP for different workloads and disk drives. Our experiments show that PCAP reduces power by up to 22%. Further, under PCAP, 60% of the requests observe service times below 100 ms compared to just 10% under throttling. PCAP also reduces worst-case latency by 50% and increases throughput by 32% relative to throttling.