The IMDEA Software researcher, Thaleia Dimitra Doudali has been awarded with a César Nombela grant. This grant will provide her with 420,000 euros for 5 years to fund the research group she is leading at the IMDEA Software Institute in Madrid, Spain. The funding is provided by the Madrid Regional Government in an effort to attract outstanding and talented young researchers from abroad and help them integrate into the Spanish research ecosystem. The research, that will be carried out with this funding, targets a critical need in industry, that is to significantly enhance efficiency and sustainability in large-scale computing environments.
Data centers and cloud computing systems consume substantial energy, accounting for 1% of the global consumption, with AI applications like generative AI and Large Language Models being particularly power-intensive. However, studies show that it has been particularly challenging to deliver high efficiency in such computing platforms, due to the suboptimal decisions of their management systems software. A critical aspect of intelligent management is the ability to accurately forecast future demand and usage of the hardware resources by the users and their applications. However, existing forecasting models, prioritizing practicality and simplicity, often overestimate resource usage, limiting efficiency gains to only 10-16% [1].
This project seeks to address this limitation by developing a more accurate resource usage prediction model, combining data-driven heuristics with only the necessary use of machine learning [2], potentially unlocking up to 60% higher resource savings compared to current solutions. At the same time, the project will explore methods to enable sustainable operations, by dynamically placing and executing workloads at times and locations that allow for reduced carbon emissions. The overall goal of the project is to deliver a resource-efficient, carbon-aware and workload-adaptive management systems software stack.
[1] Do Predictors for Resource Overcommitment Even Predict? Georgia Christofidi and Thaleia Dimitra Doudali. In Proceedings of the 4th Workshop on Machine Learning and Systems (EuroMLSys ’24). https://dl.acm.org/doi/10.1145/3642970.3655838
[2] Is Machine Learning Necessary for Cloud Resource Usage Forecasting? Georgia Christofidi, Konstantinos Papaioannou, and Thaleia Dimitra Doudali. In Proceedings of the 14th Symposium of Cloud Computing (SoCC 2023). https://dl.acm.org/doi/10.1145/3620678.3624790
