Slide 1Slide 2Slide 3BackgroundExperimental SetPower Was Estimated Using CPU UtilizationTypical Measure Result (here real data center workload)Power CappingPotential Benefits of Power CappingSpeed Scaling ExperimentsEffect of Using Speed ScalingIdle PowerSlide 13Discussion PointsPower Provisioning for a Warehouse-Size Computer(ISCA 2007)Authors: Xiabo Fan, Wolf-Dietrich Weber, and Luis Andre BarrosoGooglePresenter: Kirk Pruhs•First power usage study at data center level on real data center workloads•Findings:•First power usage study at data center level on real data center workloads•Findings:–Difference between measured peak power and aggregate theoretical per machine maximum power is 40% at data center level–Power capping can allow you to use significantly more (40%) machines without effect maximum power •More effective at data center level than rack level–Speed scaling is moderately effective (20%) at reducing average power, and has measurable effect on reducing peak powerBackground•Contracts with electrical utilities specifies peak power usage, with significant penalties for exceeding these peaks•In newest Google data centers, 85% of power hits the computing equipment –(www.google.com/corporate/green/datacenters/step2.html)Experimental Set•5K machines•Used CPU utilization to estimate power•Workloads–Websearch–Webmail–Map reduce–Mixture of the above three–Actual data center workloadPower Was Estimated Using CPU UtilizationTypical Measure Result (here real data center workload)Power Capping•Definition: Reducing power consumption to stay below some threshold–Need measurements–Option 1: Throttle the workload by delaying less essential work–Option 2: Throttle the power per machine, say by speed scalingPotential Benefits of Power CappingSpeed Scaling Experiments•When CPU utilization was below a certain threshold (5% or 20% or 50%) the contribution of CPU power was estimated to be reduced by half and the power for other components was unchanged (so there were no measurements of actual power usage)Effect of Using Speed ScalingIdle Power•First power usage study at data center level on real data center workloads•Findings:–Difference between measured peak power and aggregate theoretical per machine maximum power is 40% at data center level–Power capping can allow you to use significantly more (40%) machines without effect maximum power •More effective at data center level than rack level–Speed scaling is moderately effective (20%) at reducing average power, and has measurable effect on reducing peak powerDiscussion Points•Do we buy the conclusions reached in the paper?•How significant/interesting is the paper?•What interesting research directions does it suggest?•More specifically, what algorithmic problems does the paper
View Full Document