Roofline Model Based Performance-Aware Energy Management for Scientific Computing

Performance-aware energy management is very important for scientific computing. Inspired by the roofline model, we studied the time performance model and energy consumption model of scientific computing applications. The influence of the number of active cores and CPU frequency to the performance and energy consumption is analyzed. Based on the characteristic of computing platform and scientific computing application, the policy of determining the optimal number of active cores and frequency is proposed. Using the DVFS and power states management API, the number of active cores and their frequencies are adjusted to acquire the optimal balance between performance and energy. The proposed method has been implemented and tested with the NAS Parallel Benchmark, UGKS, and Fluent. Experimental results show that the proposed method can accurately estimate the number of optimal active cores and frequency for scientific applications and the average error of CPU power model is about 7%. Compared with the optimal performance resource allocation, the energy and performance balanced method can save 24.5% energy and just 9.8% performance degradation.