Performance Prediction for Power-Capped Applications based on Machine Learning Algorithms

Abstract

Growing high performance computing (HPC) clusters are encountering a power wall due to limitations in the surrounding infrastructure. Maximizing a cluster’s performance in the presence of a limited power budget is an open problem with high relevance and requires a deep understanding of application performance and power draw.Hardware components with the same technical specification have distinct power efficiencies and applications running on those components have diverse power profiles. Enforcing a power limit on individual components changes the performance characteristics. In this work, we investigate and quantity power- and performance-characteristics of various applications. Further, we present a systematic methodology to collect corresponding monitoring data and apply machine learning (ML) techniques to predict the performance under particular power caps. The observed prediction error is under 3% in most cases, which is in the same range of performance variation as application runs without a power cap.