BU Ph.D. candidate Fatih Acun on his innovative data center research
As artificial intelligence and supercomputing workloads continue to grow, data centers are consuming more power, putting a significant strain on the power grid. This surge in demand has created a pressing need for innovative solutions to integrate large-scale data centers with power grids.
Third-year Ph.D. student Fatih Acun is exploring ways to optimize data center power management and integrate demand response participation in an effort to solve this problem. This involves a program that encourages customers to reduce or shift their electricity use during peak periods in exchange for financial incentives, making them what he calls “flexible power consumers”. Acun’s research helps stabilize the power grids to allows sustainable growth of data centers and power grids. In November of 2024, Acun presented this groundbreaking research on GPU (Graphics processing units, integral in AI processing) power capping at the SC24 Sustainable Supercomputing Workshop.
CISE caught up with Acun to learn more about the importance of the problem at the intersection of power grids and data centers, his research towards flexibility and stability to allow sustainable growth, and his research journey.
Q: What excites you about your work on GPU power capping? Why did you want to work on this issue?
A: GPUs are the most power-demanding hardware component in data centers, and controlling their power consumption means you can control a significant percentage of the data center’s power. GPU power capping is a method to limit GPU power consumption, and it can slow down application performance. Since performance is a primary concern for large-scale computing systems, we need to carefully analyze the impact of power capping on application performance before applying power capping on production systems.
Q: Tell me about the study you’re presenting at the SC24 Sustainable Supercomputing Workshop, Analysis of Power Consumption and GPU Power Capping for MILC.
A: MILC (MIMD Lattice Computation) is a widely used supercomputing application and a lot of users submit MILC jobs with a large number of node allocations. Power consumption behavior can vary based on the application types of MILC, input size, and parallel concurrency. We performed an analysis covering various combinations of those configurations and presented our insights for GPU power consumption. Then, we analyzed the application performance under various power caps and found that there are significant energy and power-saving opportunities for MILC applications with acceptable slow-downs in their performance.
Q: What was your experience like interning with the National Energy Research Scientific Computing Center?
A: NERSC provides a really good program for summer interns. They have organized lots of seminars and activities. I also had the chance to work very closely with my supervisor Dr. Zhengji Zhao. I think it is a good opportunity for PhD students to progress in their research and get some experience by working with a top supercomputer. I was surprised about the scale when submitting jobs that allocate more than a thousand GPUs.
Q: What is the difference between working at NERSC versus in academia?
A: It is actually not too much different since it is a research center. However, they still have more strict goals because they manage a production system and they need to address the operational needs of the system and the users. In academia, I think people have a bit more flexibility to work on research ideas.
Q: What’s the one takeaway you want people to know about your research?
A: Since the power consumption of data centers increases at a very fast pace, the impact of this kind of research becomes really visible and important. Even small improvements in energy efficiency and sustainability can have significant positive outcomes for sustainable computing and power grids.
Fatih Acun received his master’s degree in 2021 from Middle East Technical University (METU) in Ankara, Turkey. While there, he worked on traffic prediction using deep learning methods. Now at Boston University, he is a third-year Ph.D. candidate studying computer engineering, advised by CISE Director and Professor Ayse Coskun (ECE, SE). In 2024, Acun was selected to be a Hariri Institute Graduate Student Fellow. Additionally, he was a research intern for the National Energy Scientific Computing Center, where he studied power capping for GPU workloads.