Power Prediction for Sustainable HPC

Abstract

Exascale computers consume huge amounts of power and their variation over time makes system energy management important. Because of time lag in cooling-units operation, predictive control is desirable for effective power control. In this work, we report a state-of-the-art power prediction model. Conventional methods with topic model use the power of past job as a prediction based on the similarity of job information. The prediction, however, fails, if there is no correct data before. To resolve this, we developed a recurrent neural network model with variable network size, which detects features of power shape from its power history and enables precise prediction during job execution. By integrating these models into a single algorithm, the optimal model is automatically adopted for prediction according to the job status. We demonstrated high-precision prediction with an average relative error of 5.7% in K computer as compared to that of 20.1% by the conventional method.