Online dynamic capacity provisioning in data centers

Abstract

Abstract-Power consumption imposes a significant cost for implementing cloud services, yet much of that power is used to maintain excess service capacity during periods of low load. In this work, we study how to avoid such waste via an online dynamic capacity provisioning. We overview recent results showing that the optimal offline algorithm for dynamic capacity provisioning has a simple structure when viewed in reverse time, and this structure can be exploited to develop a new 'lazy' online algorithm which is 3-competitive. Additionally, we analyze the performance of the more traditional approach of receding horizon control and introduce a new variant with a significantly improved worst-case performance guarantee. I. INTRODUCTIONEnergy costs represent a significant fraction of a data center's budget [1] and this fraction is expected to grow. Hence, there is a growing push to improve the energy efficiency of the data centers. A promising approach for making data centers more energy efficient is using software to dynamically 'rightsize' the data center, i.e., adapt the dispatching so that during periods of low load some servers are allowed to enter a powersaving mode (e.g., go to sleep or shut down).However, entering and leaving sleep mode incurs a penalty ("switching cost"), in terms of latency, energy consumption, or wear-and-tear. This means that decisions to sleep cannot be made independently at different time instants. The problem is challenging due to the lack of knowledge about future workloads, which means that a server that is put to sleep may soon need to be woken again. There is a significant and growing literature on this topic [2]- [5].This paper provides an overview of recent results providing online algorithms to decide the provisioning at each time instant without or with a little information of future workload. To this end, we discuss a simple but general model that captures the major issues of right-sizing. With this model, we first analytically characterize the optimal offline solution (Section III-A). We show that it exhibits a simple, 'lazy' structure when viewed in reverse time. Second, we discuss a novel, practical online algorithm motivated by this structure (Section III-B). The algorithm, named Lazy Capacity Provisioning (LCP), mimics the 'lazy' structure of the optimal algorithm, but proceeding forwards instead of backwards in time. Importantly, LCP is 3-competitive, i.e., its cost is at most 3 times that of the optimal offline solution. Third, we analyze the traditional approach of Receding Horizon Control (RHC). We show that RHC performs well when servers are homogeneous; specifically, it has performance that quickly tends toward optimality as the prediction window increases. However, we also show that RHC can perform badly when servers are heterogeneous, regardless of the length of the prediction window. To address this issue, we discuss a variant of RHC that is guaranteed to perform well in heterogeneous settings. Specifically, under both homogeneous and heterogeneous settings, their...

Cloud Computing Operations Research

“…In other cases, it is important to take randomness into account. Work in this direction has already begun; see, for instance, Al-Daoud et al (2012), Gandhi et al (2010), Lefèvre and Orgerie (2010), Lin et al (2011), Mateus and Gautam (2011), and references therein.…”

Resource Allocation: Realizing Mean-Variability-Fairness Tradeoffs

“…The choice of U V i (·) roughly corresponds to the metric proposed in [33]. To obtain a good initialization for AVR, the reward allocation in the first 10 time slots is obtained by solving a modified version of OPTAVR((m, v), c) with a simpler objective function i∈N U E i (r i ) (which does not rely on any estimates) under the same constraints (17) and (18), and run AVR from the 11th time slot initialized with parameters (m, v) set to the mean reward and half the variance in reward over the first ten time slots.…”

Resource allocation: Realizing mean-variability-fairness tradeoffs

“…Proof: We start proving the result by viewing (18)- (19) as a stochastic approximation update equation, and using Theorem 1.1 of Chapter 6 from [17] to relate (18)- (19) to the ODE (35).…”