Designing Cloud Servers for Lower Carbon Emissions

Reducing carbon emissions is crucial to curbing the effects of climate change, but usually gas-powered vehicles and manufacturers are the most conspicuous culprits. However, Information and Communication Technology (ICT) is currently responsible for between 2 and 4 percent of the global carbon footprint, which is on par with aviation emissions.

ICT emissions are expected only to grow, a byproduct of how everyday appliances like watches and cars keep getting “smarter.” ICT is predicted to be responsible for 20 percent of global carbon emissions by 2030.

Cloud computing is one of the key contributors to ICT’s carbon footprint, producing two types of emissions: operational, which arises from power generation; and embodied, which comes from the semiconductor fabs used to manufacture a cloud server’s hardware components.

Many cloud providers like Microsoft and Google have set aggressive deadlines to greatly reduce cloud emissions. With this in mind, a team of researchers from CMU identified a new design approach for compute server Stock Keeping Units (SKUs), which enable a cloud to significantly reduce carbon emissions while meeting its performance goals. The team, led by electrical and computer engineering Ph.D. student Jaylen Wang, created a systematic way to design and deploy GreenSKUs, which are carbon-efficient server designs that minimize a cloud’s overall emissions while promoting its efficiency.

In practice, the emissions vs. performance tradeoff is hard to balance.

“The amount of computing resources that an application is using is relatively proportional to its carbon emissions,” said Wang. “This means that applications that need a lot of computing power, storage, and networking capabilities will produce more carbon emissions.”

To help cloud developers compare carbon-efficient server design options, Wang and his research collaborators debuted a systematic methodology called GSF, which is notable as “the first framework for cloud providers to systematically make informed carbon-efficient server SKU design and deployment decisions.” GSF was applied in this research under Microsoft Azure’s production constraints and reduced the platform’s carbon emissions by about 10 percent. Such significant reductions in Azure’s cloud emissions can reduce 0.1%--0.2% of global carbon emissions by 2030.

Wang and the rest of his team, which includes Akshitha Sriraman, an assistant professor of electrical and computer engineering and computer science, have ideas about how to take this research further.

While the GSF framework was first tested with Microsoft, the framework is intended as a stepping stone towards enabling and accelerating the design of carbon-efficient servers across cloud computing domains, Wang said.

“The rise of machine learning and large language models like ChatGPT will make it necessary to deploy servers with specialized hardware in the cloud.”

Another reason to design cloud computing servers with more carbon-efficient components is the planned obsolescence of devices. For example, when droves of consumers upgrade their smartphones, this leads to more ICT waste and manufacturing of newer products.

“In our work we targeted planned obsolescence by reusing decommissioned components in our carbon-efficient server designs. Finding ways to use hardware longer, across computing domains, is something that this work motivates,” said Wang.