Akamas LCTES 2020 Cereda

Making AI work in performance optimization is not an easy feat. It took the Akamas team years of research to create the ML-based optimizer that powers its autonomous performance optimization platform.

But we are not stopping there: we continue to invest to make it even better, more scalable, and smarter.

Today, we’re proud to announce that our paper “A Collaborative Filtering Approach for the Automatic Tuning of Compiler Optimisations”, will be presented at the upcoming ACM International Conference on Languages, Compilers, and Tools for Embedded Systems (LCTES) taking place from June 15th. 

The paper will be presented by Stefano Cereda, an Industrial PhD Fellow at Politecnico di Milano that is part of the Akamas Research team together with Stefano Doni, Akamas CTO, and Paolo Cremonesi, Moviri’s Chief Scientific Officer and Professor at Politecnico di Milano, with the help of Gianluca Palermo, Professor at Politecnico di Milano. Stefano’s Ph.D. research focuses on the automatic tuning of IT systems’ configurations. 

The Industrial Fellowship program is a great example of Moviri’s and Akamas’ productive collaboration with the school of engineering at Politecnico di Milano. This partnership continues to generate applied innovations at the nexus of technology and its application to business problems.

The LCTES conference is a premier scientific forum linking the software programming and embedded systems engineering communities. By exposing researchers and developers from their respective domains to relevant research in programming languages, algorithms, and AI, the conference stimulates cross-pollination and the advancement of research and industry. 

In this paper, we propose a novel approach to optimize compiler flags by using recommender system techniques (such as the ones used in technologies like ContentWise’s UX Engine). Contrary to what other research suggested, we show how traditional characterization techniques based on workload metrics can actually mislead the compiler auto-tuning task. We present our new methodology based on Collaborative Filtering techniques, and we show it outperforms the current state of the art approaches.