Data Center Power Cost Optimization via Workload Modulation

Abstract

Abstract-We formulate optimization problems to study how data centers might modulate their power demands for cost-effective operation taking into account various complexities exhibited by real-world electricity pricing schemes. For computational tractability reasons, we work with a fluid model for power demands which we imagine can be modulated using two abstract knobs of demand dropping and demand delaying (each with its associated penalties or costs). We consider both stochastically known and completely unknown inputs, which are likely to capture different data center scenarios. Using empirical evaluation with both realworld and synthetic power demands and real-world prices, we demonstrate the efficacy of our techniques. I. INTRODUCTIONIt is well-known that data centers incur significant costs towards powering and cooling their computing, networking, and storage (IT) equipment, making the optimization of these costs an important research area. Most electric utilities employ one (or a combination) of the following pricing mechanisms: (i) peak-based pricing, wherein a component of the electricity bill is dependent on the peak power drawn over the billing cycle (typically a month) [2], (ii) time-varying pricing, wherein the per unit energy price fluctuates over time (there are many examples of this operating at different time granularity such as very fine timescale spot prices [11] or higher prices during periods of "coincident peaks" experienced by the utility [7]), and (iii) load-dependent or tiered pricing, wherein higher energy prices are applied for higher energy consumption [9].Whereas recent research has begun looking at data center cost optimization bearing some of these pricing complexities in mind, this is still a nascent field in many ways. In particular, most instances of such work (e.g., [4]) focus exclusively on a particular pricing feature, whereas real-world tariffs often combine these features, making cost-effective data center operation even more complicated. This is the context of our paper: how should data centers optimize their costs given the various features of real-world electricity pricing?Broadly speaking, the vast literature on data center power cost optimization may be understood as employing one or both of (i) demand-side (based on modulating the data center's power demand from within) and/or (ii) supply-side (based on employing additional sources of energy supply the data center's existing utility provider and backup sources) techniques. Our focus is on techniques based on "IT knobs," a subset of (i) that relies

A Hierarchical Demand Response Framework for Data Center Power Cost Optimization under Real-World Electricity Pricing

“…We omit it here for space and present it in a technical report [25]. Discussion: We point out some limitations and possible generalizations of our formulations below:…”

“…Similar notations can be defined to handle the uncertainties in electricity price; we omit modeling electricity price as random variables for simplicity and present the full version with electricity price in [25].…”

“…This specialized SDP (that we label SDP Drop ) has a reduced runtime of O(R · L 3 p · T ). More interestingly, the structure of the optimal policy for SDP Drop given by the following lemma (proof in technical report [25]) provides a way of explicitly finding the optimal control policy. Define…”

“…Furthermore, ON Drop offers a competitive ratio of 2 − 1 n under peak-based pricing when α t = α for all t. We present our proofs in [25]. One useful way of understanding ON Drop is to view it as a generalization of the classic ski-rental problem [11]: increasing the dropping threshold is analogous to purchasing skis while admitting/dropping demands according to the existing threshold is analogous to renting them.…”

“…As described in Section II, since several prior works have looked at DR for time-varying utility prices, we restrict the focus of our evaluation to peak-based pricing (while noting that our framework is general enough to also capture timevarying prices including coincident pricing as in [14]). We present our experimental evaluation with time-varying utility pricing in our technical report [25].…”