GDCEngine: an end-to-end machine learning engine for green data center control optimization with digital twins

Abstract

The ever-increasing scales and power consumption of data centers (DCs) have brought challenges that aim to minimize energy costs while avoiding operational risk. Current industry practice mainly relies on the data centre infrastructure management (DCIM) system and requires extensive human expertise, which is difficult to scale with the growing DC complexity. To advance DC management, we propose GDCEngine, an end-to-end green data center (GDC) AI engine that facilities the development and applications of machine learning (ML) approaches for DC optimizations. In GDCEngine, we develop three major components to support the training and evaluation of ML-based policies for different user groups. First, we develop the GDCSimulator module that integrates multiphysics digital twin simulation for evaluating ML-based policies without risking the physical system. Second, we develop the GDCPolicy module that incorporates advanced deep reinforcement learning algorithms for DC cooling control optimizations. Third, we design a no-code web interface to facilitate the usage of DC operators without intensive prior knowledge in ML. We use a case study to demonstrate the developed engine in optimizing a chilled water cooling DC based on our proposed Safari algorithms. The experimental evaluation shows that GDCEngine helps save 26% total power consumption compared with a conventional controller, and dramatically reduces thermal safety violations during the online learning stage. GDCEngine is a firm step towards transforming GDC ML research and application.