Effective Use of Performance Monitoring Counters for Run-Time Prediction of Power

Power dissipation and energy consumption have recently become first-order design constraints for microprocessors. In this paper we present the measured power dissipation of several SPEC benchmarks executing on a Pentium 4 processor. Using on-chip performance monitoring counters, we dynamically correlate relevant performance metrics such as instructions per cycle, second-level cache hits, and bogus/non-bogus instructions retired to the measured power. This quantitative analysis lends insight into selecting metrics to be used for predicting power. We found that incorrectly predicted branches had the strongest correlation, of 0.8979, to power dissipation, while execution time had the highest correlation to total energy consumption, 0.99.