EDP Optimization of Parallel Applications via CPU Frequency Scaling on AMD Processors
Mariana C. Toledo, Sandro M. Marques, Thiarles S. Medeiros, Fábio Diniz Rossi, Marcelo Caggiani Luizelli, Antonio Carlos Schneider Beck, Arthur F. Lorenzon · 2021
Dynamic Voltage and Frequency Scaling (DVFS) has been used to improve the use of computational resources when a system is executing parallel applications. On top of that, as parallel applications have different behavior, DVFS methods must be able to deal with the application characteristics at hand. However, as we show in this paper, DVFS governors already available in many Linux Operating System distributions are not capable of dealing with such a scenario, providing a tradeoff between performance and energy consumption (energy-delay product - EDP) that is distant from the best possible. Given that, we propose a run-time and dynamic approach to optimize the EDP of parallel applications running on AMD processors that automatically selects the ideal CPU frequency and Boosting operating mode according to the characteristics of the application at hand. When executing sixteen well-known parallel applications on two multicore architectures, we show that our approach provides EDP optimizations of up 38% when compared to the ondemand DVFS governor.