We introduce the EP-queue -a significant generalization of the M B /G/1 queue that has state-dependent service time probability distributions and incorporates power-up for first arrivals and power-down for idle periods. We derive exact results for the busy time and response time distributions. From these we derive power consumption metrics during non-idle periods and overall response time metrics, which together provide a single measure of the trade-off between energy and performance. We illustrate these trade-offs for some policies and show how numerical results can provide insights into system behavior. The EP-queue has application to storage systems, especially hard disks, and other data center components such as compute servers, networking and even hyper-converged infrastructure.
Introduction
Trading off power and response time in modern computing infrastructuresModern software-defined data centers are typically organized in a hierarchical fashion, with multiple tiers of cache present within and across components ranging from processors to network and storage components. Advances in caching technologies (e.g. host-side flash caches) mean that much read traffic is being absorbed before it reaches the lower tiers [11,15]. Thus the workloads of lower tiers are increasingly dominated by writes which can be potentially buffered in higher tiers. Nevertheless, occasional read bursts still occur due to the reading of as-yet-uncached data.This burstiness of access creates the possibility of using components that can be turned off during idle periods and restarted when new work arrives, thus saving energy while delivering acceptable access latencies. Efficient usage of compute, network and storage resources is important in light of the fact that data centers are collectively storing about 35-50% more data per year [1] and consuming more than one percent of global electricity [9]. Much progress has been made on the server side with application consolidation via server virtualization and shared networked storage.However, further energy optimizations will be required as long as idle sub-components consume energy.In a system with components that shut down to save power, we need to understand the expected busy period durations to determine power savings. Similarly, we need to incorporate startup (POWER UP) and shutdown (POWER DOWN) times as random variables to model the impact on both the power and response time characteristics. As one would expect, the first arrival to an idle server will incur a significant delay but subsequent arrivals will likely fare much better. We would like to characterize the impact of this kind of response time behavior on host-side applications. It is in this context that we introduce and analyze the Energy-Performance (EP) queue -a significant generalization of the M B /G/1 queue that has state-dependent service time probability distributions and incorporates power-up for first arrivals and power-down for idle periods. The analytical model developed compares two basic policies: the two-...