A comparison of high-level full-system power models
HotPower'08 Proceedings of the 2008 conference on Power aware computing and systems
Chaotic attractor prediction for server run-time energy consumption
HotPower'10 Proceedings of the 2010 international conference on Power aware computing and systems
ACM Transactions on Architecture and Code Optimization (TACO)
Towards energy-proportional computing for enterprise-class server workloads
Proceedings of the 4th ACM/SPEC International Conference on Performance Engineering
Hi-index | 0.00 |
Energy efficiency is an important concern in computer systems from small handheld devices to large data centers and supercomputers. Improving energy efficiency requires metrics and models: metrics to assess designs and identify promising energy-efficient technologies, and models to understand the effects of resource utilization decisions on power consumption. To facilitate energy-efficiency improvements, this dissertation presents JouleSort, the first completely specified full-system energy-efficiency benchmark; and Mantis, a generic and portable approach to real-time, full-system power modeling. JouleSort was the first full-system energy-efficiency benchmark with fully specified workload, metric, and rules. This dissertation describes the benchmark design, highlighting the challenges and pitfalls of energy-efficiency benchmarking that distinguish it from benchmarking pure performance. It also describes the design of the machine with the highest known JouleSort score. This machine, consisting of a commodity mobile CPU and 13 laptop drives connected by server-style I/O interfaces, differs greatly from today's commercially available servers. Mantis generates full-system power models by correlating AC power measurements with software utilization metrics. This dissertation will evaluate several different families of Mantis-generated models on several computer systems with widely varying components and power footprints, identifying models that are both highly accurate and highly portable. This evaluation demonstrates the trade-off between simplicity and accuracy, and it also shows the limitations of previously proposed models based solely on OS-reported component utilization. The simplicity of this black-box approach makes it a useful tool for power-aware scheduling and analysis.