Mantis and JouleSort: Power Models and Metrics in Enterprise Computing

Опубликовано 6 сентября 2016, 16:46

Improving the energy efficiency of enterprise computing requires new ways of designing systems and new policies for adaptively managing individual systems or data centers. To enable these improvements, this talk presents Mantis, a non-intrusive method of modeling full- system power consumption and providing real-time power predictions; and JouleSort, the first complete full-system energy-efficiency benchmark. Mantis generates models by correlating AC power measurements with high-level system utilization metrics. We experimentally validate the models on two server systems with drastically different power footprints and characteristics, using a variety of workloads. Mantis provides power estimates with high accuracy for both average and instantaneous power consumption, making it a useful tool for power- aware scheduling and analysis. JouleSort is the first complete full-system energy-efficiency benchmark, with fully specified workload, metric, and rules. It is a variant of the popular Sort Benchmark, which has historically been used to evaluate system performance and price-performance. This talk will motivate the benchmark, discuss its design (including some subtle pitfalls), and describe the machines we designed that achieve the highest benchmark scores.