Supercomputing and Sustainability: Reducing the Carbon Footprint of Big Data Centers

When researchers perform extensive computations with large amounts of data, they send their algorithms and data to supercomputing centers that house hundreds of servers that perform extensive parallel computing and process massive amounts of data. These computing systems have led astrophysicists to gain a better understanding of the universe, enabled epidemiologists to track daily rates of Covid-19, and made it possible for climatologists to better understand historic weather patterns. Yet these supercomputing centers consume a large amount of energy. Energy is consumed as the server cores process data, as communication happens between servers, as coolant systems keep the machine racks cool (as all of this processing and communication generates heat), and as central air conditioning systems run continuously to keep the centers cool. Because of the high energy consumption of these centers, big data comes with a big carbon footprint.

Ram Ramesh, a professor of management science and systems at the University at Buffalo in the United States has worked with two other researchers to analyze how these supercomputing centers might reduce their carbon footprint.

Inside a supercomputing system with many cores

There are two processes that use energy during data analysis, computing and communication. “Each job requires computing over a set of cores requested by the user and this is not a process that a data centre administrator can change,” says Ramesh. This means the energy used for the actual computing cannot be reduced. The other process that uses energy is communication across cores.

Communication from server to server is much more energy expensive than communication within a server. So, when possible keeping jobs together in the same server would reduce energy consumption, and a packed server produces a lot of heat. Cooling things down also takes energy. Ramesh and his colleagues were interested in devising an algorithm that would help a supercomputer administrator to know when it was most efficient to pack servers with jobs (to put jobs together within a small set of servers) and when to balance the load (to distribute the work across the servers).

Through their data collection at the supercomputing center in Singapore, the researchers came to understand the threshold levels that lead to the most energy efficient computing. If the total core requirement of all jobs is above a certain threshold, it is more energy efficient to pack loads into as few servers as possible. If the total core load of jobs is below the threshold, then it is more energy efficient to balance the loads across servers.

Ramesh and his colleagues’ research shows a 10-30% overall energy reduction when this system is used. This system is “a viable practical solution that allows an administrator to easily make effective decisions for energy conservation,” says Ramesh. The dissemination of this research to other supercomputing centers could reduce the carbon footprint of supercomputing centers globally. Their work appears in the leading academic journal Information Systems Research.

You can find out more about the supercomputers that Dr Ramesh directly works with and on here

Ramesh and his colleagues, Zhiling Guo and Jin Li, conducted their research in a large supercomputing center in Singapore. Housing 20 racks of servers, each rack holds 72 servers and each server contains 24 cores. Ramesh explains that each job sent to the computing center may require several concurrently running cores, and these cores can be dispersed across servers both within and across racks. This means the data center administrator needs to decide at any time how to allocate cores to requesting jobs across the 34,560 cores (20 racks x 72 servers x 24 cores). Where the administrator sends those jobs has an effect on how much energy is used. Yet, typically the speed of job completion is the only factor in that decision, not energy conservation.