pHPCe: a hybrid power conservation approach for containerized HPC environment

Reducing power consumption with tolerable performance degradation is a fundamental challenge in today's containerized High Performance Computing (HPC). Most Dynamic Power Management (DPM) approaches proposed for the HPC environment are based on profile-guided power-performance prediction techniques. However, the complexity of DPM approaches in a multi-tenant containerized HPC environment (cHPCe) increases significantly due to the varying demands of users and the contention of shared resources. Moreover, there is limited research into software-level monitoring of power consumption in the popular Docker container environment since it is not designed keeping HPC in mind. The proposed research in this paper aims to present a real-time hybrid power-performance prediction approach using Long Short-Term Memory (LSTM) machine learning model. Unlike state-of-the-art techniques, LSTM with the rolling update mechanism accurately predicts the relationships between the tail latencies in time-series power performance data. It updates the training sample sequences according to the currently predicted power consumption and duration to predict sporadic power surges. It also proposes a power-cap determination framework with resource contention awareness to fine-tune power consumption in real time at the thread level. The proposed containerized environment is designed from scratch keeping HPC requirements in mind. Hence, we name the proposed approach as power-aware cHPCe (pcHPCe) and is evaluated and compared with native BareMetal execution using the NAS Parallel Benchmark (NPB) and HPC Challenge benchmark (HPCC) applications. Our experimental results show that the power-performance prediction model achieves an accuracy of 91.39% on average in real-time with an overhead of 1.6% of the total computing power per node. Our resource contention-aware power-cap selection framework attains significant power saving up to 13.1%.