A Quantitative Evaluation of the RAPL Power Control System

We evaluate Intel’s RAPL power control system, which allows users to set a power limit and then tunes processor behavior to re-spect that limit. We evaluate RAPL by setting power limits and running a number of standard benchmarks. We quantify RAPL along five metrics: stability, accuracy, settling time, overshoot, and efficiency. The first four are standard measures for evaluating con-trol systems. The last recognizes that any power control approach should deliver the highest possible performance achievable within the power limit. Our results show that RAPL performs well on the four standard metrics, but some benchmarks fail to achieve maxi-mum performance. At high power limits, the average performance is within 90 % of optimal. At middle power limits, it is 86 % of op-timal. At low power limits, the average performance is less than 65 % of optimal. 1