The computational demands of science, industry and artificial intelligence are growing rapidly. One way to reduce both processing time and energy consumption is software autotuning – a method that allows software to automatically search for the most efficient way to run. Jiří Filipovič from Masaryk University’s Institute of Computer Science has now succeeded as the main proposer of a new COST Action that will connect European experts in autotuning. In this interview, he explains why existing tools need to be better aligned, what the network can offer early-career researchers, and how computer game development first led him into science.
Our work has two closely connected directions. The first is collaboration on the development of scientific software. For example, computational biologists come to us with a particular problem, their own program, or an idea for a solution, but the computation is not fast enough.
First, we need to understand the problem itself. Then we look for ways to improve the computation – from the mathematical design all the way down to implementation details – so that more data can be processed or results obtained faster.
The second direction is the development of general methods for speeding up software. We focus especially on autotuning, which allows a program to automatically search for the settings under which it will run most efficiently.
The two directions complement each other. Practical collaborations show us what autotuning tools need to be able to do if they are to work in real applications. We then use the results of our research when developing specific scientific software
At CERIT-SC, we partly act as advanced support for users of the computing infrastructure. We can look at the computations of a particular research group, advise them on how to run them more efficiently, or become directly involved in developing the software they need.
CERIT-SC also allows us to test methods on real-world tasks and collaborate with experts from other scientific disciplines. At the same time, research into autotuning extends beyond CERIT-SC and Masaryk University itself. We develop general approaches that other groups can use on different machines and in completely different applications.
“We can look at the computations of a particular research group, advise them on how to run them more efficiently, or become directly involved in developing the software they need.”
The aim is to connect the European community working on autotuning. Today, many research groups are developing different tools and frameworks. These often operate at different levels, have difficulty exchanging data, and use different methods to evaluate results. As a result, two papers may address a very similar problem, but their results cannot easily be compared because each uses a different methodology.
The COST Action should create a space for joint work on standards, tool integration and evaluation methods. Another goal is to make autotuning accessible to a broader group of users. Advanced tools today often require changes to source code and are intended mainly for experts. We therefore want to connect autotuning specialists with developers of high-level programming languages.
A user could, for example, simply mark which part of a program may be modified, while the tool would automatically search for the most suitable variant. Artificial intelligence also has considerable potential here.
“The COST Action should create a space for joint work on standards, tool integration and evaluation methods.”
A COST Action does not primarily fund the research itself or staff positions. It mainly supports the development of an international network. It can fund working-group meetings, workshops, summer schools, conferences, short-term stays at institutions abroad, or the development of educational materials. The idea is to bring together people working on similar problems who would not normally have the opportunity to collaborate.
Some frameworks, for example, modify the computing environment; others adjust hardware settings; others modify the source code. To connect them, experts need to agree on how the individual optimization steps should interact: what can be tuned separately, what needs to be tuned together, in what order, and how data should be passed between different autotuning layers. That is the kind of problem one research group can only address with difficulty on its own.
“The idea is to bring together people working on similar problems who would not normally have the opportunity to collaborate.”
One of the first steps will be a summer school designed to attract young researchers and expand the community. We will also begin harmonizing individual autotuning frameworks. Some tools already exchange data to a certain extent, but deeper integration requires collaboration among several teams.
Short-term scientific missions will also be important. An autotuning expert might, for example, visit a group developing weather forecasting software and help them integrate our technologies. Similarly, we can invite one of their researchers to the summer school.
The goal is to bring these methods into specific scientific applications. At the same time, broader collaboration among different groups will allow us to explore more fundamental research questions. For example, how software should decide when to launch autotuning, how many resources to allocate to tuning, or at which levels tuning should take place.
“Broader collaboration among different groups will allow us to explore more fundamental research questions.”
It was mainly a long-term process. You need to build a network of contacts and establish professional credibility so that others are willing to work with you on a project and accept you as the lead proposer.
First, we had to assemble a group of people willing to participate and then prepare a coherent proposal, even though several partners were contributing to it.
With large projects, it sometimes happens that everyone adds their own piece, and the result resembles a cake made by too many cooks – everything is in there, but together it does not really make sense. This time, we had enough time to bring the individual parts together and give the whole proposal a shared logic.
It helped that we are addressing a very current problem. The computational demands of science, industry, and artificial intelligence are growing rapidly. At the same time, energy consumption and the environmental footprint of computing systems are increasing.
Autotuning offers a way to make software faster while reducing the energy required to run it. Even a ten-percent speed-up can be significant if an institution spends tens or hundreds of millions of Czech crowns on computing infrastructure.
I think the evaluators appreciated the combination of a timely topic, a well-constructed international network, and a proposal that clearly described the problems we want to solve.
“Autotuning offers a way to make software faster while reducing the energy required to run it. Even a ten-percent speed-up can be significant if an institution spends tens or hundreds of millions of Czech crowns on computing infrastructure.”
If the Institute of Computer Science organizes and coordinates a major international network, it strengthens its professional reputation.
Even more important, however, are the new contacts. Meetings, workshops, and scientific stays will connect people and help them become familiar with our work. In the future, that should make it easier to prepare joint European projects, for example, within Horizon Europe.
PhD students and early-career researchers will also gain major opportunities. They will be able to take part in research stays and summer schools, and work at institutions abroad. The network can therefore generate collaborations, research topics, and projects that may only emerge in the years ahead.
“Meetings, workshops, and scientific stays will connect people and help them become familiar with our work.”
I actually got into research through games. I enjoyed creating physical models and optimizing them to run as fast as possible. From there, it was a direct path to computationally demanding problems and software optimization. The experience from game development also comes back to me when building a community.
When we were students and worked as independent game developers, we started organizing the Game Developers Session conference. It was originally meant simply as an opportunity for us to meet each other, but over time, professionals from major studios began attending as well.
I was the main organizer of the conference from its inception in 2003 until 2008, after which I moved into academia full-time. So even back then, I was already helping build a community and creating opportunities for people to meet. Now – after quite a long break – I am beginning to do something similar internationally in the field of science. :-)
Perhaps less than one might think. In both cases, you have to persuade people to do something without being paid for it, reply on time, and meet deadlines. In the end, it is almost irrelevant whether you are organizing a group of beer-drinking university students or busy scientists. 😊 But I enjoy this kind of work. I am the sort of person who is happy to shut myself away for a week to focus on one problem, but I also like bringing people together.
I am curious myself whether, four years from now, I will regret having submitted the proposal, because leading the network will mean a lot of work. On the other hand, at a certain stage of one’s career, it is also natural to move beyond influencing only your own research and to start shaping the environment around you.
Jiří Filipovič
works at the Institute of Computer Science of Masaryk University, where he leads a research group focused on High Performance Computing. His research primarily focuses on automatic software performance optimisation, autotuning, GPU acceleration, and the development of computationally intensive applications for scientific use, including applications in computational chemistry and biology.
He also serves as a supervisor and lecturer at the Faculty of Informatics, Masaryk University. In 2021, he received the MUNI Scientist Award for outstanding research achievements.
A COST Action is a four-year international research network supported by COST – European Cooperation in Science and Technology, bringing together researchers, innovators and other experts across Europe. Unlike conventional grants, COST Actions do not primarily fund research itself. Instead, they support collaboration through activities such as workshops, summer schools, short-term scientific missions and other networking activities. Support is awarded through open calls, with selected proposals undergoing independent expert evaluation. Success therefore represents not only an opportunity to build an international network, but also significant professional recognition for the proposer and the participating institutions.
Jiří Filipovič will lead the new COST Action AutoPEAC, bringing together European experts in software autotuning, performance optimization, and energy efficiency.
Generative AI is no longer just an experiment for enthusiasts at Masaryk University. Students use it in their studies, teachers in preparing their courses, and researchers when working with texts, data and code. Most of us are already familiar with Microsoft Copilot, an assistant integrated into the Microsoft 365 environment.
