Achieving informed carbon reduction in the data center

Earlier this year, the United Nations Framework Convention on Climate Change (UNFCCC) reported that current climate commitments will not sufficiently limit global warming. To restrict global temperature increases to 1.5C, we all need to do more. Arm has committed to achieve carbon net-zero by 2030 by 2030 and is working to decarbonize compute, leveraging low-power, high-performance foundational technology to drive down energy consumption and help reduce emissions.

In pursuit of this, my team and I investigated how we can operationalize the decarbonization of compute beyond the surface level for ourselves and our partners. Our research formed part of our participation in the UN Global Compact’s Young SDG Innovators Program (YSIP). The 10-month program sees teams of young professionals develop solutions that deliver market value for their company but also drive progress on the Sustainable Development Goals.

The result was a tool called the Carbon Advisor (CA), which allows users to make informed decisions about the carbon impact of their compute workloads. By considering carbon alongside cost, performance and quality it helps users choose between cloud providers, instance types, and compute locations.

The carbon cost of datacenters

So why did we focus on datacenters?

In 2010, datacenters consumed around 1% of global electricity use. But the share of total compute electricity use is shifting towards the cloud and hyperscale, efficiency improvements are becoming harder and harder to attain, and demand is increasing. We need a data-oriented approach to achieve a clear view of the carbon cost of our activities, so we can use that information to lower carbon emissions.

Estimated global data electricity use by data center type, 2010 and 2018.            Source: Masanet et al (2020). Energy innovation: How Much Energy Do Data Centers Really Use?	Recalibrating global data center energy-use estimates. Source: Masanet, Eric & Shehabi, Arman & Lei, Nuoa & Smith, Sarah & Koomey, Jonathan. (2020). Science. 367. 984-986. 10.1126/science.aba3758.

When assessing how best to design, maintain, and utilize datacenter compute – and minimize the greenhouse gas (GHG) emissions of a datacenter – there are some factors to consider:

• Grid carbon intensity
  The amount of carbon from energy use, typically expressed in grams of CO₂ equivalent per kilowatt-hour of energy. Other GHGs have an associated Global Warming Potential (GWP) value that is converted to a carbon equivalent.
  
  
• Embodied carbon emissions
  The greenhouse gas emissions arising from the manufacturing, transportation, installation, maintenance, and disposal of building materials and equipment.
  
  
• Facility space utilization
  The use of existing IT equipment relative to the total available space for IT. Effective space utilization can ease cooling costs and lower overall carbon emissions.
  
  
• Server hardware power efficiency
  Design power efficiency inherent to the servers used in the datacenter. Many factors can affect this, including processor type, workload efficiency, and server age distribution.
  
  
• Power Usage Effectiveness (PUE)
  The facility energy overhead as a ratio of the IT load. A high ratio indicates higher energy need and thus a gap in efficiency that increases the total carbon emissions of a facility for its utility.

This is not a complete list, and yet it demonstrates the wide variety of influences that affect the final carbon cost of a datacenter. Some effects are determined at the earliest stages in a datacenter’s development, like the manufacturing of silicon contributing to embodied carbon emissions. Others only come into play during active operation, such as grid carbon intensity. There is also a variety of data availability represented: grid carbon intensity statistics are typically easier to procure than embodied emissions, for instance, and only recently have approaches been attempted that measure holistically across this larger life cycle.

An operationalizing solution must start somewhere. Consideration data availability and the influence that Arm is most capable of (a topic covered later on), our team focused primarily on applying measurements of grid carbon intensity and server hardware power efficiency.

The carbon advisor

The tool shows data views from both the datacenter user (carbon emissions per product) and provider (performance output per product) perspectives and supports comparison of up to three solutions. By selecting parameters such as the target workload, compute source and type, and location, users can gain a breakdown of the carbon emissions associated with each solution. Using this information, users can make more carbon-conscious decisions about their resource management with a finer granularity and full awareness of their final carbon impact. Rarely is carbon cost considered a primary design decision; the Carbon Advisor is a step towards enabling that practical consideration.

An anonymized screenshot of the prototype Arm Carbon Advisor

Our first task was to identify the differing requirements of each user group.

Defining user requirements

A user wants to accomplish a set level of work emitting as little carbon as possible. They can achieve this by choosing their cloud provider, type and size of instance, and where their resources are housed as mindfully as possible to limit their emissions. Shown below are the factors they control directly or indirectly in this process.

Defining provider requirements

The provider aims to maximize the total performance/utility for a given workload within the bounds of a fixed power (and carbon) amount.

The data challenge

Achieving this level of transparency and granularity in applying the carbon lens to the datacenter is feasible, but procuring reliable and transparent data is a challenge. The following examples show the complexity of the issue:

• Grid Carbon Intensity: The carbon intensity of the electricity supplied to a datacenter has a profound impact on its final emissions. While the electricity grid of a city, state, or country often have publicly available carbon metrics, many cloud providers procure renewable energy for their datacenters that considerably decrease the carbon intensity of their power. These numbers are often difficult to procure, but are necessary to make accurate calculations for any given solution.
  
  
• Server Power Draw: The different processors used within a datacenters' servers have different power-draw requirements. Thermal Design Power (TDP) is often used as a proxy for power-draw metrics for benchmarking and estimations, but in a datacenter context TDP is often inaccurate. This is due to the power utilization and design of server racks, as well as other common methods used by datacenters, like virtualization. Procuring accurate numbers requires detailed Total Cost of Ownership (TCO) analysis in a server rack context, which is not always possible.
  
  
• Workload performance: In practice, carbon cost will always be balanced with the need to retain adequate performance. However, making an informed tradeoff requires accurate performance estimates for target workloads. There are common workloads for which most processors have publicly available estimates, like SPEC Integer Rate. But for many workloads this information is not readily available, and costly to ascertain independently.

Current standards of data availability, accuracy, and transparency limits access to concrete data on carbon metrics. Without that concrete data, making informed carbon-conscious decisions is a difficult task. The CA makes the caveats and assumptions made in its calculations clear, but an ideal solution would utilize data that is fully vetted and transparent. This is something that can only be possible through open communication between vendors and users. To achieve the larger sustainability goals of the ICT industry – and the world – a concerted and collaborative approach across our industry is needed.

[CTAToken URL = "https://www.arm.com/sustainability" target="_blank" text="Learn more about Sustainability at Arm" class ="green"]