MIT researcher applies explainable AI to data center efficiency
Machine learning for more efficient cloud infrastructure
MIT associate professor Christina Delimitrou is applying machine learning to make large-scale data centers more efficient, secure and reliable. Her work addresses a growing environmental concern: power-hungry data centers can strain electrical grids and increase dependence on polluting fossil-fuel energy sources.
Delimitrou, the KDD Career Development Associate Professor in Communications and Technology in MIT’s Department of Electrical Engineering and Computer Science and a member of the Computer Science and Artificial Intelligence Laboratory, is rethinking how servers, networks and cloud software operate. The aim is to obtain more computing capacity from hardware that data-center operators already own.
Her research group redesigns outdated cloud systems, develops approaches to manage shared hardware resources, and creates streamlined server architectures. Delimitrou argues that underused infrastructure consumes more power than necessary to meet user demand. Removing software bloat without compromising performance could reduce the need to build as many new facilities.
From low utilization to AI-assisted operations
During work at Stanford University, Delimitrou and her research mentor Christos Kozyrakis found that many large computing systems were running at only about 15 percent capacity, rather than close to full utilization. The gap made resource efficiency and sustainable scaling central technical problems for her subsequent work.
Machine learning can automate cloud resource-management decisions at a scale that is difficult for developers to handle manually. Delimitrou says empirical approaches require substantial expertise, while the size of these systems can make it hard for operators to identify effective actions on their own.
One tool developed by her group, Seer, uses deep learning to anticipate and prevent web-application problems before they occur. It is designed to avert broad slowdowns that can follow a manual fix. The research also responds to the shift toward applications split into smaller components across multiple servers, a design approach that speeds deployment but does not necessarily match the capabilities of existing servers.
Explainability and realistic testing
At MIT, Delimitrou’s group uses AI to redesign software systems around available hardware and has extended debugging research beyond code errors to security issues that can expose user data. The team is also working to make AI tools explainable, because an uninterpretable recommendation can limit a developer’s ability to verify an answer or learn how to improve a system.
Research access is another constraint. Data-center operators increasingly rely on proprietary hardware and software that academic teams cannot directly examine. To study these environments, Delimitrou’s group developed Ditto, a tool that mimics an application’s structure and performance characteristics so researchers can conduct a wider range of experiments.
Delimitrou expects more capable machine-learning models to create additional opportunities to improve application performance and hardware efficiency, while stressing that AI must be audited carefully. For businesses operating cloud services, the practical implication is to assess utilization, software overhead and failure handling alongside new hardware purchases, and to retain human review of AI-driven operational decisions.