How sustainable is the cloud, really?

Measuring and addressing emissions in data centers and the cloud.

Photo by Lucy Chian on Unsplash

I’m guessing data centers aren’t the first thing that comes to mind when you think of sustainability. Data centers and data transmission networks consume 1–1.5% of all electricity worldwide and 0.6% of total GHG emissions. I’ve discussed previously how data center emissions map to Scope 1, 2, and 3 emissions, but for most internal IT departments or CIO offices the type of emissions will separate out along the poles of private on-prem data centers, colocation data centers, and public cloud services.

Clients are responsible for emissions tied to different hardware and software layers depending on the operational model.

On-prem data centers

On-prem data centers emissions have Scope 1 emissions produced by their heating and cooling infrastructure and Scope 2 emissions from purchased electricity. Efficiency at this level is measured through several metrics:

• Total energy consumption: Measured in kWh and typically captured for both IT equipment alone and the total data center including overhead and cooling.
• PUE: Power usage effectiveness is a metric of data center energy efficiency. It is the ratio of energy consumed by IT equipment compared to the total overhead energy consumed. For example, a PUE value of 1.5 uses an additional 0.5 kWh of overhead energy for every 1 kWh consumed by IT equipment. A “perfect” PUE value is 1.0 and most data centers aim to be below 1.4. While useful, PUE is a flawed metric because it can be artificially improved by increasing IT equipment energy consumption in ways that are not sustainable such as overprovisioning hardware.
• CUE: Carbon usage effectiveness is different than PUE in that it has a unit (kg CO2 per kWh). It is the ratio of total CO2 emissions (carbon footprint) from the total data center energy consumption compared to the energy consumed by IT equipment.
• WUE: Water usage effectiveness is the ratio of total liters of water used in the data center for cooling and humidity control to the energy consumed by the IT equipment. Average WUE is 1.8 L per kWh.
• Percent renewable energy: Renewable energy delivered to the grid is factored into the regional carbon emission factor (CEF). Additionally, many data center providers purchase renewable energy certificates (RECs) through power purchase agreements (PPAs) to claim a higher percentage of renewable energy. There are different factors to consider when purchasing RECs. Some companies have strict requirements on the types of RECs they will accept so it’s important to be familiar with your company’s standards.

To calculate these metrics, utility meter readings and bills are the main source of total energy and water consumption data for on-prem data centers. IT equipment energy consumption is best measured via intelligent power distribution units (PDUs) and power monitoring software like PowerIQ. For older IT equipment lacking PDUs, energy consumption can be estimated by multiplying the nameplate power value on each server in kW by the total number of hours the device is run (typically 8,760 hours per year) at 50% utilization. CUE is calculated by multiplying PUE by the Carbon Emission Factor (CEF), which is dependent on the energy source and geography. There are several sources for CEF estimated values including the EPA for US locations and the IEA for global locations. Many companies are starting to rely on sustainability software, such as Envizi and others, for calculating their enterprise carbon footprint including data centers.

86% of companies, are addressing emissions from on-prem data centers by migrating to the cloud. Carbon emissions can be reduced by an estimated 70–90% by moving workloads to the cloud, enabling greater asset utilization, carbon aware workload management, and access to renewable energy sources. If your IT org still maintains on-prem data centers, check out the Green Grid organization for examples of how to drive overhead efficiency, and many of the recommendations below still apply.

Colocation data centers

Emissions from these cloud models fall under Scope 3. The metrics above still apply and should be considered when you select a provider and manage workloads, but customers have less control over the factors that impact them. Today, many landlords share these metrics in monthly or annual customer reports. Colo customers also frequently use intelligent PDUs described above to monitor daily energy usage. If your landlord does not yet disclose this data or leverage renewable energy credits, you can use your buying power to encourage them to do so.

To address Scope 3 emissions tied to colos, IT organizations should be aware that they control IT equipment energy consumption and therefore directly impact PUE, CUE, and WUE. There are several ways to reduce these emissions including:

• Consider replacing dated servers with newer, more energy efficient models.
• Optimize server resource allocation with application resource management tools, which the IBM CIO is leveraging. Resources are frequently overprovisioned to maximize reliability, driving up the need for additional hardware. Optimizing server utilization and using techniques like serverless computing can allow you to potentially turn servers off or at least slow the demand for new hardware. Observability tools can also provide awareness, enabling application performance improvements while reducing cost and energy consumption
• A related, but potentially harder problem is detecting and decommissioning zombie servers or underutilized hardware, which can consume as much as 60% of a fully utilized server. Today this is a largely manual process. Regular audits can help, but automated tools are on the horizon, including work coming out of IBM Research.
• Containerizing workloads can reduce annual infrastructure costs by up to 75%. Containers can be easily spun up when needed and also enable greater portability across architectures and machines.
• Although still in early days, carbon-aware workload scheduling holds a lot of promise. In its simplest form, this can be choosing to run workloads at times of lower energy demand, avoiding power from “peaker plants” that are typically run on fossil fuels. One recent example is Microsoft scheduling updates for when local carbon intensity is lowest. Carbon awareness can also mean selecting geolocations with reduced carbon intensity. Some tools already exist to automate this process.

“As a service” models

Many major cloud service providers enable access to monthly energy consumption by service, location, and cost center through dashboards. There is also the open source Cloud Carbon Footprint, which uses cloud provider usage data, PUE, and regional carbon intensity to calculate the carbon footprint of workloads in Azure, AWS, and Google clouds. In general, emissions tied to “as a Service” models like PaaS and Saas can be addressed by practicing sustainable architecture and development practices. Beyond that, Microsoft is starting to explore carbon aware scaling. Google has been implementing carbon aware location selection for certain media processing workloads and also provides region picker to allow their cloud customers to balance cost, latency, and carbon footprint.

At the end of the day, balancing customer need for performant cloud computing resources with sustainability is the goal, and that requires first baselining our key metrics and setting meaningful targets to improve them. There’s a ton of innovation in this space, and I’m excited to see what comes next. Definitely share any useful tools or practices you’ve encountered in your sustainable cloud journey.

Stacey Gifford, Ph.D. is CIO Sustainability Lead at IBM based in New York. This article is personal and does not necessarily represent IBM’s positions, strategies, or opinions.