The amount of CO $_{2}$ emitted per kilowatt-hour on an electricity grid varies by time of day and substantially varies by location due to the types of generation. Networked collections of warehouse scale computers, sometimes called Hyperscale Computing, emit more carbon than needed if operated without regard to these variations in carbon intensity. This paper introduces Google’s system for global Carbon-Intelligent Compute Management, which actively minimizes electricity-based carbon footprint and power infrastructure costs by delaying temporally flexible workloads. The core component of the system is a suite of analytical pipelines used to gather the next day’s carbon intensity forecasts, train day-ahead demand prediction models, and use risk-aware optimization to generate the next day’s carbon-aware Virtual Capacity Curves (VCCs) for all datacenter clusters across Google’s fleet. VCCs impose hourly limits on resources available to temporally flexible workloads while preserving overall daily capacity, enabling all such workloads to complete within a day with high probability. Data from Google’s in-production operation shows that VCCs effectively limit hourly capacity when the grid’s energy supply mix is carbon intensive and delay the execution of temporally flexible workloads to “greener” times.

Over the past decade, there has been a global growth in datacenter capacity, power consumption and the associated costs. Accurate mapping of datacenter resource usage (CPU, RAM, etc.) and hardware configurations (servers, accelerators, etc.) to its power consumption is necessary for efficient long-term infrastructure planning and real-time compute load management. This paper presents two types of statistical power models that relate CPU usage of Google’s Power Distribution Units (PDUs, commonly referred to as power domains) to their power consumption. The models are deployed in production and are used for cost- and carbon-aware load management, power provisioning and infrastructure rightsizing. They are simple, interpretable and exhibit uniformly high prediction accuracy in modeling power domains with large diversity of hardware configurations and workload types across Google fleet. A multi-year validation of the deployed models demonstrate that they can predict power with less than 5% Mean Absolute Percent Error (MAPE) for more than 95% diverse PDUs across Google fleet. This performance matches the best reported accuracies coming from studies that focus on specific workload types, hardware platforms and, typically, more complex statistical models.

Despite impressive recent advances in reinforcement learning (RL), its deployment in real-world physical systems is often complicated by unexpected events, limited data, and the potential for expensive failures. In this paper, we describe an application of RL “in the wild” to the task of regulating temperatures and airflow inside a large-scale data center (DC). Adopting a data-driven, model-based approach, we demonstrate that an RL agent with little prior knowledge is able to effectively and safely regulate conditions on a server floor after just a few hours of exploration, while improving operational efficiency relative to existing PID controllers.

Safely exploring an unknown dynamical system is critical to the deployment of reinforcement learning (RL) in physical systems where failures may have catastrophic consequences. In scenarios where one knows little about the dynamics, diverse transition data covering relevant regions of state-action space is needed to apply either model-based or model-free RL. Motivated by the cooling of Google's data centers, we study how one can safely identify the parameters of a system model with a desired accuracy and confidence level. In particular, we focus on learning an unknown linear system with Gaussian noise assuming only that, initially, a nominal safe action is known. Define safety as satisfying specific linear constraints on the state space (e.g., requirements on process variable) that must hold over the span of an entire trajectory, and given a Probably Approximately Correct (PAC) style bound on the estimation error of model parameters, we show how to compute safe regions of action space by gradually growing a ball around the nominal safe action. One can apply any exploration strategy where actions are chosen from such safe regions. Experiments on a stylized model of data center cooling dynamics show how computing proper safe regions can increase the sample efficiency of safe exploration.