Evaluating the Advantage of Reactive MPI-aware Power Control Policies

Power consumption is an essential factor that worsens the performance and costs of today and future supercomputer installations. In state-of-the-art works, some approaches have been proposed to reduce the energy consumption of scientific applications by reducing the operating frequency of the computational elements during MPI communication regions. State-of-the-art algorithms rely on the capability of predicting at execution time the duration of these communication regions before their execution. The COUNTDOWN approach tries to do the same by mean of a purely reactive timer based policy. In this paper, we compare the COUNTDOWN algorithm with state-of-the-art predictive-based algorithm, showing that timer based policies are more effective in extract power saving opportunities and reducing energy waste with a lower overhead. When running in a Tier1 system, COUNTDOWN achieves 5% more energy saving with lower overhead than state-of-the-art proactive policy. This suggests that reactive policies are more suited then proactive approaches for communication-aware power management algorithms.