Making AI Systems More Adaptive by Professor Peter Pietzuch

Abstract

Distributed AI model training and inference are becoming the most important workloads in modern data centres. Despite the fast progress in the domain of AI systems, we still do not know the best way to design a distributed AI software stack. In this talk, I will trace the evolution of AI software stacks and make the case that existing designs are not sufficiently adaptive -- they fail to react adequately to changes in data centre resources and workload requirements at runtime. This poses new interesting systems research challenges on how to design the next generation of adaptive AI systems. Based on our work at Imperial College London, I will give examples of new abstractions that can add adaptive features to today's AI software stacks with minimal changes and low overhead.


Biography
Peter Pietzuch is a Professor of Distributed Systems at Imperial College London, where he leads the Large-scale Data & Systems (LSDS) group (https://lsds.doc.ic.ac.uk). His work focuses on the design and engineering of scalable, reliable and secure data-intensive systems, with a particular interest in machine learning, data management and cloud computing systems. Currently, he serves as a Programme Committee Co-Chair for the ACM European Conference on Computer Systems (ACM EuroSys 2026) and a General Co-Chair for the International Conference on Very Large Data Bases (VLDB) 2025. Until recently, he was the Director of Research in the Department of Computing, a Co-Director for Imperial's I-X initiative on AI, data and digital and the Chair of the ACM SIGOPS European Chapter~(EuroSys). In 2023, he received the ACM SIGMOD Test-of-Time Award for his work on scalable stream processing systems. Before joining Imperial College London, he was a post-doctoral Fellow at Harvard University. He holds PhD and MA degrees from the University of Cambridge.