Offline performance and energy consumption prediction model of deep learning training tasks

Abstract

Today, deep learning technology is widely used in a lot of scientific research fields. A crucial question is how to accurately predict the performance and energy consumption of deep learning training (DLT) tasks. Existing prediction methods of DLT tasks either have low accuracy or use too many cluster resources, and few methods focus on the energy consumption prediction. In this paper, we analyze the relationships between the characteristics of performance and energy consumption and the task configurations of DLT tasks. Then we propose an offline prediction model to predict the performance and energy consumption of DLT tasks based on these relationships. The experiment in an actual GPU cluster shows the effectiveness of the prediction model. The average deviation of the prediction model is 4.68%.