Forecasting power-efficiency related key performance indicators for modern data centers using LSTMs

Modern HPC and cloud data centers, being mission critical facilities, continue to grow in size and number as the amount of Internet and cloud-based services increases. Moreover, the recent hype in the usage of machine learning based technologies, requiring appropriate infrastructures for data storage and computational power, further pushes the need for modern data centers. This expenditure not only increases the operational costs, but also affects the environment by generating carbon footprints. Due to increased load on supporting power grids, some governmental organizations start to reconsider data center deployment procedures with an increased demand of renewable energy utilization and waste heat recovery. One of the challenges when powering data centers using renewable energy is the intermittent availability of the power. The dynamics of available power must be taken into account to maximize usage of accessible renewable energy while adhering to existing service level agreements as well as to thermal and power requirements. This requires a framework that can continuously match the actual behavior of the target computing infrastructure to the current availability of power and cooling. This paper introduces machine learning based approaches aiming to model various data center energy/power consumption related Key Performance Indicators (KPIs). A validation is performed using the multi-year operational data obtained at Leibniz Supercomputing Centre (LRZ). This framework is used as a building block for achieving a data center infrastructure-aware resource management and scheduling. (c) 2020 Elsevier B.V. All rights reserved.