Carbon-Aware Computing for Datacenters

Abstract

The amount of CO2 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.

Power Modeling for Effective Datacenter Planning and Compute Management

“…The PDU-level CPU usage of serving jobs (category 1) in Google datacenters have predictable daily patterns. On the other hand, the total PDU-level CPU usage of batch processes has considerable intraday variability and their intraday profiles are hard to accurately predict [9]. As a result, the total PDU-level CPU usage typically varies significantly within a day with fluctuations larger than 10% for more than half of the PDUs across the Google fleet.…”

“…1) Per-PDU Power Model: To enable power efficiency and carbon-aware workload management as in [9], we use a Extensive analysis of 5-minute average PDU power consumption u PDU POW , as a function of its average CPU usage u PDU CPU , indicates three distinct utilization regimes (Figure 3). Furthermore, PDU power consumption monotonically increases with its CPU usage.…”

“…The Per-PDU model's piecewise linearity provides a suitable framework to evaluate the impact of changes in CPU usage on power usage, which can then be easily integrated into cost-and carbon-aware optimizations designed to shape compute load by shifting compute tasks in time and space ( [9], [17], [18]), i.e., between datacenter clusters. However, while the discussed models have comparable MAPEs when predicting changes in power usage as a result of small variations in CPU usage, this is not the case when changes in CPU Cluster level performance of the proposed models for eight experimental clusters during the planned power drop event.…”

“…During summer months of 2021, Test II was conducted to evaluate Google datacenters potential to drop power for 6 hours on selected days: 7/13, 7/29, 7/30, 7/31, 8/11. The power load was dropped by enforcing lower cluster-level CPU consumption using the capacity curve mechanism that leverages temporal flexibility of some of compute workloads, as recently proposed in [9]. The load shedding was accomplished in two stages lasting 2 and 4 hours to avoid negative performance implications on workloads due to the abrupt and significant drops in available CPU capacity.…”

“…Each model shows advantages in certain use cases. For example, the Per-PDU model's predictive performance across all CPU/power usage regimes and its piecewise-linearity is exploited within Google's Carbon-Intelligent Computing System [9] to evaluate the impact of changes in datacenter clusters' CPU usage on their power consumption as a result of shifting workloads in time and across locations. The Per-PDU model has also been used for estimation and real-life validation of a datacenter's potential to reduce power consumption during grid-level demand response events.…”