Design and Evaluation of an Energy Agile Computing Cluster

AbstractVariable, intermittent renewable sources of energy arebeing introduced into the national electric grid at scale.However, the current grid paradigm of load-followingsupplies is unable to utilize these sources without im-practical deployments of backup generation and grid-scale energy storage. Thus, it will be crucial to haveenergy agile loads: electric loads that can dynamicallyadapt their energy consumption. Data centers are primecandidates for becoming energy agile because they arelarge energy consumers and often have workloads thatcan be shaped to vary their power consumption. In thispaper we present an energy agile cluster that is powerproportional (uses power proportional to its workload)and exposes slack (the ability to temporarily delay ordegrade service to reduce power consumption). We de-scribe a prototype cluster that consumes 60% less en-ergy on typical workloads by being power proportional.Then, using month-long traces from a 9572 node com-puting cluster and a California wind farm, we show howa grid-aware scheduler can use workload slack to reducesdependence on non-renewable energy sources to 40% ofits original level. Our results show that achieving thesame wind penetration with energy storage alone wouldrequire sufficient battery capacity to run a cluster for fivehours.