Advanced Computing in the Age of AI | Sunday, July 21, 2024

Toward a Low-Carbon Datacenter 

According to a new study, large datacenters could slash greenhouse gas emissions by 88 percent by switching to highly efficient equipment and implementing energy management practices. 

IT has an important role to play in reducing carbon emissions; one of the surest ways to reduce energy consumption is to make the technology itself more efficient. Renewable energy sources are another integral part of a sustainable energy strategy, but equipment/infrastructure inefficiencies are wasteful of clean and dirty sources alike. This is the principle finding of a recent article in the esteemed journal Nature Climate Change, authored by green energy researchers Eric Masanet, Arman Shehabi and Jonathan Koomey.

The report is titled "Characteristics of low-carbon data centres." Why datacenters? Well the authors assert that 1) the global population of datacenters use 1-2 percent of the world's electricity and are responsible for about 0.5 percent of carbon emissions; and 2) these infrastructures have lots of inefficiencies.

So there is still unmet opportunity for global datacenters as the backbone of the Internet to be greener, but this is not an indictment against the "dirty datacenter." In fact, a growing body of evidence suggests that the Internet has reduced societal carbon emissions by replacing carbon-intensive physical mediums (like books, cd's, magazines and most notably mail) with their digital equivalents. Furthermore, Internet- or cloud-based services are likely to be more efficient than local computations.

For its part, the IT industry keeps finding ways to increase energy efficiency and has begun to select for lower-carbon energy sources as well, but the authors argue that there are economic and environmental motivations for accelerating this transition.

To this end, the researchers have identified three areas for improvement (which Koomey reprises in a blog article):

1. The efficiency of the information technology equipment, i.e., servers, storage, and communications.

2. The efficiency of the infrastructure equipment, i.e. fans, cooling, pumps, power distribution.

3. The carbon intensity of electricity production.

A study mentioned in the Nature report notes that on average, a large US datacenter could reduce its carbon emissions by 88 percent by switching to state-of-the-art equipment and implementing energy management practices.

Because datacenters contribute to greenhouse gas emissions in known ways, the problem can be attacked systematically starting with the actions that will have the most effect. Maximizing IT device efficiency goes hand-in-hand with other measures, like high-level utilization. Location can have a huge impact on the overall carbon profile of a datacenter; free cooling and/or low-carbon energy sources are the main considerations.

While decisions to purchase new equipment should take into account embedded carbon costs, the researchers maintain that operational efficiency and carbon performance carry more weight in the low-carbon datacenter. Based on their work, the researchers conclude that "embodied emissions are likely to be small on an absolute basis and the fastest path to reducing such emissions is reducing device counts through operational efficiency."

Then there's the issue of measurement. As the authors write, the current "imperfect and evolving portfolio of performance metrics obscures which data centre characteristics correspond to low-carbon operations." For industry and government stakeholders, the authors highlight the potential for energy models to provide actionable guidance.

An energy-carbon performance map plots operational energy use (y axis) against electric power CO2 intensity (x axis). It's a fascinating exercise that handily illustrates greenhouse gas emissions as being dependent on technology, location and electricity source characteristics.

In the words of the researchers, "the map also highlights an important distinction: although the use of renewable electricity can lead to significantly better carbon performance, it does nothing to improve energy performance. Indeed, low-carbon electricity without low-energy operations is suboptimal because, given their high energy intensity, data centres now using renewable power almost exclusively draw from centralized and resource-constrained supplies (notably hydropower, wind, geothermal and biogas). In other words, high-energy data centres unnecessarily tie up 'low-carbon' electrons that might otherwise be used to reduce emissions from other customers in a region."

This key point is distilled down to one sentence in the report's final section:

Although renewable energy can significantly reduce a data centre's carbon emissions, an inefficient (that is, high-energy) data centre will use far more low-carbon electricity than is technically required.

Koomey, writing on his blog, asserts that while infrastructure efficiency is important, IT efficiency should be the primary aim:

IT efficiency (which includes higher utilization and performance improvements as well as purchasing efficient hardware) is the most important issue on which to focus. Most recent efforts in the industry have been on improving infrastructure efficiency, which has many beneficial effects, but is not as important a lever as is the IT efficiency (in many new facilities we are reaching the limits of infrastructure efficiency, with PUEs as low as 1.04).

The map also drives home the limitations of the Power Usage Effectiveness (PUE) metric. Developed by The Green Grid, a datacenter's PUE is calculated by dividing its total electricity use by the electricity consumed by the IT equipment. (Lower scores are more desirable with 1.0 being the ideal score.) When device efficiency is low, reducing PUE can have a nice effect on carbon footprint. However, when IT device efficiency is at the upper limits, PUE is less influential.

To restate this disconnect in real-world terms, a datacenter with a mediocre PUE of 1.8 may be greener (with regard to energy and carbon) than a datacenter with a best-in-class PUE of 1.1. (For a clear example, compare points 2D and 3D on the map above.) The PUE has become a de facto industry standard when it was never intended to be used for cross-facility comparisons. However, as the authors of the Nature article advise, "PUE and electricity source can play a role in assessing carbon performance when used concurrently."