Power Modeling for Effective Datacenter Planning and Compute Management

Abstract

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.

Carbon-Aware Computing for Datacenters

“…Our estimate of this upper bound, used for resource reservations, is greater than usage with high probability (close to 1). In prior work [20], we demonstrated that the power usage of a cluster power domain can be estimated within 5% error using a piecewise linear function of its CPU usage. As a result, any change in cluster-level CPU usage can be accurately mapped into a change in its power usage, which is critical for building CICS (see Subsection III-A for more details).…”

“…One item that differentiates this new methodology from previous approaches is its inclusion of risk associated with application and infrastructure performance expectations. Special attention is given to (i) predicting the next day's flexible and inflexible compute usage, (ii) translating it to power consumption [20], and (iii) then optimizing in a risk-aware manner. To meet Google's infrastructure and application SLOs, there is monitoring, performance tracking, and a feedback loop that evaluates the recent application level impact and controls load shaping accordingly.…”

“…Figure 2 shows a simplified view of power architecture of a typical Google datacenter [20], [47]. Every datacenter is connected to the electricity grid via several medium voltage feeders.…”

“…The aggregate cluster-level load shaping is achieved using a Virtual Capacity Curve (VCC), which artificially limits the cluster's hourly compute usage and, in view of the power modeling advancements in [20], its hourly power usage as well. Borg uses the VCC values to compute the real-time CPU availability for incoming flexible jobs.…”

“…• First-of-its-kind demonstration: To the best of our knowledge, this is the first demonstration of carbon-aware algorithms that shifts datacenter computing in time to realize global environmental and efficiency objectives [19], using automated adjustments based on current and forecasted grid conditions. • Effective, risk-aware, optimization approach: Our load shaping mechanism uses a risk-aware optimization approach that incorporates fleetwide, cluster-level load forecasts, power models [20], explicit workload and infrastructure performance expectations, as well as datacenter power-contract information. The demonstrated effectiveness of the proposed mechanism is due to the high accuracy of the cluster-level load forecasts and power models.…”