Our research is currently supported by the following projects.
- NSF CAREER: A Robust and Data-driven Design for Carbon-intelligent Distributed Systems, 2021-2026, PI: Mohammad Hajiesmaili
- NSF Collaborative Research: CNS Core: Medium: Dynamic Data-driven Systems – Theory and Applications, 2021-2024, PI team: Mohammad Hajiesmaili, Ramesh Sitaraman (UMass Amherst), Steven Low, Adam Wierman (Caltech), Zhenhua Liu (Stony Brook University)
- NSF Collaborative Research: CNS Core: Small: Dynamic Pricing and Procurement for Distributed Networked Platforms, 2021-2024, PI team: Mohammad Hajiesmaili (UMass Amherst), Carlee Joe-Wong (CMU)
- NSF/VMWare Collaborative Research: NGSDI: CarbonFirst: A Sustainable and Reliable Carbon-Centric Cloud-Edge Software Infrastructure, 2021-2024, PI Team: Mohammad Hajiesmaili, David Irwin, Ramesh Sitaraman, Prashant Shenoy (UMass Amherst), Adam Wierman (Caltech), Tian Guo (WPI)
- NSF Collaborative Research: CPS Medium: Enabling DER Integration via Redesign of Information Flows, 2021-2024, PI team: Mohammad Hajiesmaili, Prashant Shenoy (UMass Amherst), Dennice Gayme, Enrique Mallada (JHU), Steven Low, Adam Wierman (Caltech)
NSF CNS: Core: Small: Energy and Load Management in Data Centers: Online Optimization and Learning, 2019-2022, PI: Mohammad Hajiesmaili
In recent years, there has been an unprecedented growth in energy footprint and cost of data centers as critical infrastructures of Internet services. Powering data centers from a portfolio of energy sources and elasticity in load execution are two key features that enable optimizing the energy cost by shifting the energy load in the temporal domain. This project intends to develop energy cost minimization algorithms that utilize the potentials of shifting the data center energy load in time. Having extreme and multidimensional uncertainty as the fundamental challenge, this proposal proposes a disciplined research based on algorithmic and learning understandings of the underlying optimization problems. Successful implementation of this proposal provides a beneficial design that can optimize the cost and robustness of data center energy operations and hence will have a significant impact on lowering the overall cost of Internet services.
This project focuses on three key goals of data center energy operations: minimizing the energy cost, maximizing the robustness against uncertainty, and improving the energy footprint of data centers. Toward this, it underpins the theoretical foundations of energy and load management in data centers from online algorithm and learning perspectives. Both are promising approaches since they do not rely on any exact or stochastic modeling of the future, hence, they are robust against extreme variability. The cornerstone of this proposal is to decompose the general problem into two subproblems of energy procurement and load management. Then, it develops robust and cost-effective online algorithms for each subproblem with provable performance guarantees against severe uncertainty. Finally, by leveraging the insights from the algorithm design for the subproblems, and the optimal offline solutions, it develops efficient learning algorithms for the general problem. This project is divided into three major tasks:
- Online Optimization for Energy Procurement: A disciplined design that tackles the cost minimization data center energy procurement. It designs online algorithms with provably best performance guarantees.
- Online Optimization for Load Management: It designs cost-effective and deadline-aware algorithms for managing elastic loads in data centers, with provable competitiveness against optimal offline solutions.
- Online Learning for Joint Energy and Load Management: A learning-based framework that learns joint optimization of the energy procurement and scheduling of the elastic load.
More broadly, this research brings fundamental algorithmic and learning understandings into data center energy research and also extends the existing results on classic online conversion problems, i.e., the online search for best prices in order to buy and/or sell assets, with single-dimensional uncertainty, to the multidimensional uncertainty. Hence, the proposed research makes fundamental theoretical contributions by advancing conversion problems to address multidimensional uncertainty. Finally, data centers are key infrastructures that enable a variety of Internet services. Hence, our work on optimizing the cost of a data center will have a significant positive impact on lowering the cost of Internet services. More broadly, it will facilitate the efficient and reliable incorporation of renewables into the data center, thereby reducing the carbon footprint of data centers. Such improvements will play a key role in moving toward a more sustainable data center and the electric grid.
Lin Yang, Mohammad H. Hajiesmaili, Ramesh Sitaraman, Adam Wierman, Enrique Mallada, and Wing Wong, ‘‘Online Linear Optimization with Inventory Management Constraints,’’ in Proc. of ACM Sigmetrics, Boston, MA, USA, 2020, to appear.
Sohaib Ahmad, Arielle Rosenthal, Mohammad H. Hajiesmaili, and Ramesh Sitaraman, ‘‘Learning from Optimal: Energy Procurement Strategies for Data Centers,’’ in Proc. of ACM eEnergy, Phoenix, AR, USA, 2019.
- Russell Lee, Mohammad H. Hajiesmaili, and Jian Li, “ Learning-Assisted Competitive Algorithms for Peak-Aware Energy Scheduling,” preprint.