Managing Power Consumption and Performance of Computing Systems Using Reinforcement Learning (Poster ID T8) G. Tesauro, R. Das, H. Chan, J. Kephart, C. Lefurgy, D. Levine, F. Rawson: IBM Research Utility Managing power consumption is a major issue in data centers Major new application domain for RL: we learn policy for dynamic control of HTTP CPU clock speed trading application Req. performance vs. consumed power We employ a realistic test-bed We demonstrate utility gain over hand-designed policies & HTTP obvious RL implementations Wkld. CPU Frequency Controller IBM Blade IBM Blade RL ?Power & ?Perf. IBM Blade Power (watts) Perf. (msec) 0 Time (x5 sec) 4K