Energy-Efficient Dynamic Provisioning in Data CentersJoint work with Tan Lu from The Chinese University of Hong Kong, Lachlan Andrew from Swinburne University of Technology, and Ramesh K. Sitaraman from University of Massachusetts Amherst
Energy consumption represents a significant cost in data center operation. In 2010, data centers world-wide consumed 240 billion kWh electricity (1.3% of the world total), enough to power 5 Hong Kong or roughly the entire Spain. However, real-world statistics reveals that a large fraction of the energy is used to power idle servers when the workload is low. Dynamic provisioning techniques aim at saving this portion of the energy, by turning off unnecessary servers. In dynamic provisioning, it is a common approach to predict future workload to certain extent and exploit the information to achieve good performance. This naturally leads to the following fundamental questions:
In this work, we seek answers to the above questions. In particular, we develop online dynamic provisioning solutions with and without future workload information available. We first reveal an elegant structure of the off-line dynamic provisioning problem, which allows us to characterize the optimal solution in a “divide-and-conquer” manner. We then exploit this insight to design two online algorithms with competitive ratios 2-alpha and e/(e-1+alpha) , respectively, where 0<= alpha <=1 is the normalized size of a look-ahead window in which exact workload prediction is available. We prove that these competitive ratios are the best possible for deterministic and randomized algorithms; hence, they characterize the benefit of predicting future workload. A fundamental observation is that future workload information beyond the full-size look-ahead window (corresponding to alpha=1) will not improve dynamic provisioning performance. We remark that our results hold as long as the overall energy demands (including mainly server, cooling, and power conditioning) is a convex and increasing function in the total number of active servers. Our algorithms are decentralized and easy to implement. We demonstrate up to 71% of energy saving in a case study using real-world traces. Publications
Presentation
|