ECE PhD Prospectus Defense: Fatih Acun

ECE PhD Prospectus Defense: Fatih Acun

Title: From Load to Grid Asset: Collaborative Methods for Enabling Multi-Participant Data Center Demand Response

Presenter: Fatih Acun

Advisor: Professor Ayse K. Coskun

Chair: Professor Emiliano Dall'Anese

Committee: Professor Ayse K. Coskun, Professor Ioannis Ch. Paschalidis, Professor Emiliano Dall'Anese, Dr. Carole-Jean Wu

Google scholar Profile: https://scholar.google.com/citations?hl=en&user=iwSBZeYAAAAJ&view_op=list_works&sortby=pubdate

Abstract: The rapid growth of data centers has raised substantial concerns due to their high energy demands, particularly with the surge in artificial intelligence workloads. As large-scale power consumers, data centers put significant stress on power grids, driving up the demand for expanded generation and transmission capacity, as well as increasing electricity costs. To address grid stability and match power supply and demand, Independent System Operators (ISOs) offer demand response (DR) programs to utilize the power consumption flexibility of demand-side consumers and allow resilient growth of power grids. Although data centers are well-positioned to participate in DR through power management and workload scheduling techniques, real-world adoption remains limited, partially due to concerns over workload performance degradation and compatibility with DR program requirements.

This thesis argues that collaborative DR participation among data centers (or data center tenants) can meet quality-of-service (QoS) requirements while maximizing benefits through coordinated flexibility across participants. Towards this claim, the thesis makes three contributions at the intersection of data center energy management and power grids: (1) design of collaborative frameworks for multi-participant data center DR, using the aggregate flexibility to mitigate the risks of violating QoS requirements and improve DR power tracking capabilities, (2) design and open-source implementation of a data center DR simulator to enable testing of power management policies at scale for various DR programs, (3) analysis of the power consumption and performance outcomes of power management methods of workloads on GPU-based production level large-scale systems.

The overarching goal of this thesis is to develop methods that enhance the performance of data centers concerning QoS constraints and DR requirements, while enabling their operation as flexible power consumers to relieve the strain on power grids. To this end, this thesis approaches the problem from both theoretical and practical perspectives by employing simulation and collaborative optimization techniques, as well as empirical studies of power management on production systems. Our results show that, using collaborative methods, data centers can mitigate QoS and power tracking violations while reducing energy costs by up to 11.2% compared to individual DR participation. Future work includes the deployment of collaborative frameworks on real systems to demonstrate DR participation on GPU clusters and the design of predictive models to capture uncertainties in data center DR participation.