Servers use a lot of energy in data centers, but what many don’t realize is that their environmental impact starts even before they’re placed in massive, air-conditioned buildings. The process of gathering raw materials and building parts like processors, motherboards, and memory modules also uses energy and contributes to carbon emissions.
To address this, researchers from Carnegie Mellon University, Microsoft, and the University of Washington are working on a project to reduce these emissions. They’ve developed a prototype server called GreenSKUs, which is designed to run on Microsoft’s Azure cloud platform. The term "SKU" stands for Stock Keeping Unit, a way the tech industry labels products.
In their research, presented at the International Symposium on Computer Architecture, the team introduced a new method to identify older server parts that are still usable, reducing waste.
They found that many servers are discarded just because one part is not functioning well, even though other parts still have plenty of life left. Ashkitha Sriraman, a professor at Carnegie Mellon, explained that it's common for an entire server to be thrown away when only one component fails. The researchers are also improving the GreenSKUs by adding software that helps decide which tasks should be run on the refurbished servers versus newer ones. This helps make sure the GreenSKUs perform well without sacrificing energy efficiency.
The GreenSKUs use parts like RAM modules and solid-state drives from old, non-working servers. These reused components, paired with a more energy-efficient processor, help lower the carbon emissions that would otherwise come from running these servers.
Cutting Carbon Emissions in Cloud Computing
As cloud computing grows, reducing its carbon emissions is becoming more important. The Association for Computing Machinery’s Technology Policy Council predicts that by 2030, cloud computing could account for 20% of global emissions. Currently, the cloud industry already uses about 3% of the world’s total energy each year. However, there is a clear path to reducing these emissions. Cloud companies are focused on using less energy and increasing their use of renewable energy sources.
For example, researchers looked at Microsoft Azure’s cloud operations and found that by reusing older server parts and improving energy efficiency, they could reduce both the energy used and the emissions created. This method cut the total carbon emissions—both from building the servers and running them—by 8%.
When applied on a larger scale, the researchers estimate that this approach could reduce global carbon emissions by 0.1 to 0.2%. While this might sound like a small number, it's actually significant. It's similar to the amount of carbon emissions produced by all smartphone use in the United States.
Making Server Reuse Possible with New Technology
Currently, Microsoft replaces its Azure servers every 3 to 5 years to keep them running efficiently, but many parts that could still be used in other servers are discarded. To make it easier to reuse parts, the researchers decided to take advantage of newer technology that allows for backward compatibility, meaning older parts can still work with newer systems.
For example, Compute Express Link (CXL) controllers can now connect processors, memory, and drives, and they support the two latest types of memory modules, DDR4 and DDR5.
To help server engineers reuse parts without causing performance issues, the researchers created a system that helps identify which used components can be reused safely. This system makes sure the reused parts won't slow down performance or use so much extra energy that it defeats the purpose of reusing them.
Solving the Challenges of Using Older Components
The researchers found that reusing older components came with some challenges, and they had to come up with solutions to make it work.
The older RAM caused some delays and slower memory speeds, but the team used a technique called Pond’s approach to improve performance. Similarly, older SSDs had slower speeds and lower bandwidth, but they solved this by using a method called RAID striping, which helped boost performance.
Ashkitha Sriraman explained that while they’re saving on carbon emissions from manufacturing new parts, they’re not sacrificing too much in terms of performance.
The AMD Bergamo processors used for better energy efficiency had lower speeds and smaller caches than the processors typically used in Azure servers. There wasn’t a perfect solution for this issue, but the researchers included this limitation in their overall system, helping them match the right combination of parts to meet performance needs.
In the end, the research shows that you don’t always need the newest technology to meet the demands of cloud services. As Sriraman pointed out, “Many of these applications work just fine with older components.”
