Interview with Peter Düben, ECMWF, Head of Earth System Modelling
Well hello, Peter. Let's talk about Mantik.
The "ML Development Infrastructure Tool", as it is officially called in our research project MAELSTROM!
An exciting EU research project that aims to use machine learning on supercomputers for weather and climate forecasting! But you were reported to have said about Mantik: "It won't work anyway."
No, that's not true! – wait, is that even the official part of the interview? – When we applied for funding for MAELSTROM in 2018, we had a plan - and the plan was good: In order to be able to contrive applications, we needed a workbench that everyone could use, that people could use to communicate and start the runs, that offered flexibility and allowed optimization. The central problem was that we had to develop applications and dev tools at the same time. This means that the application developers can't work with the software from day 1 because it simply isn't there yet. But the idea is still good and completely valid.
You said "we had a good plan". Against what background was this plan written? What was the world like back then?
In our field – weather and climate – people were starting to work with machine learning. But we weren't experts yet. We had to learn a sh…load about neural networks and how to prepare the data. But also: How can you train software effectively and know how good it is at any point during training? How can you compare different methods? How can you send something to the supercomputer that you previously developed on a laptop?
But machine learning wasn't new at the time, even if it wasn't yet so common for weather and climate. Was there nothing from other areas back then you could use to get started with minor adaptations?
It was no longer new, but it was still in a fledgling stage. If you look at what was done with ML in 2018 and what is done today, there are worlds apart. That doesn't just apply to weather and climate, but to all applications. What MAELSTROM deals with, high-performance computing and supercomputing, has changed fundamentally. You've had exponential growth since 2018. But a lot has also happened in the software sector since 2018. Some tools have emerged that already do a lot of what we were planning back then, such as ML Flow or Weights & Biases.
How far has Mantik come now? And will it continue to be relevant, or has it been caught pants-down and overtaken by others?
The tool idea is still good and relevant, and I'm now looking forward to seeing how all the threads are brought together over the next six months. The first MAELSTROM applications will also be running in Mantik at the end of September. Then you will see exactly what you can do with the tool. And then, if Mantik can shine, I assume that more and more users will become interested in it. When we get out of this beta tester phase, where it's always "this will be here next week, this could work soon...", we can really cross the Rubicon.
Supercomputing is a special topic of yours. What else makes the field of weather and climate special, in terms of requirements for a tool?
We have very, very large data sets – ginormous samples. Our efficiency depends very much on how we handle data. Our data is simply more particular than cat videos or photos of horses. Otherwise, the requirements are relatively similar. When I hear that Mantik is used a lot in the automotive industry, I can well imagine because large data sets are involved there, too. And there you probably have 500 developers who have to work well in concert.
Are you saying this because collaboration is a central feature of Mantik?
Yes. You can collaborate via GitHub repositories, and if you have a question, you can ask via Slack. But Mantik is another level: you can compare and benchmark what different developers are doing.
What is the recipe for success that ultimately makes a tool become a standard? Just like GitHub is a kind of standard. Regardless of whether it's machine learning or HPC or whatever?
It has to solve a problem. Git, for example, solves the problem of how to distribute code and bring together the work of ten different people. And Mantik could link everything together quite a tad better: you endow Git repositories with more interfaces, facilitate communication between developers, but also offer the possibility to start and train runs from its GUI. The strength of Mantik is that it connects the individual packages with each other. Whether that's ML Flow or GitHub. It's not an easy task because the individual components are dynamic and evolve.
That sounds like a huge task.
At any rate, it's not a task that will be over any time soon.
Tell me a bit about Markus Abel - you've probably had the most to do with him. What's it like working with him? How does he approach things?
Markus is an entrepreneur. He can sell what he does. He's not afraid to take big steps, which is very important in a venture like this. And so it's a lot of fun to work with him. Markus is also open to criticism. He has lots of great ideas. He's sometimes a bit over-optimistic, but I think that's also part of his job description.
It's interesting that you say he's an entrepreneur. Some other people I ask would say he's still deep in ...
... the university.
The university and research.
Not from my perspective. Probably because my own head is still pretty much stuck in the research sandbox.
What is the best-case outcome if you combine the two worlds - traditional research and entrepreneurship?
It's a healthy process all round. You can also see this from the fact that, as far as I know, Ambrosys has never had any problems finding young talent. What works is that you pick people up at university and then nurture them in a small to medium-sized company.
Why do you think that is?
It's definitely also down to people like Markus, who on the one hand have an open ear and on the other aren't afraid to hire someone who doesn't have 30 years of experience in the trade.
Tell us something critical about Markus and his people.
I've already said everything I can think of: Markus is sometimes a bit overoptimistic. But that's part of his job description. There's really nothing upsetting.
To whom or for which tasks would you recommend Ambrosys?
For all situations like ours back then: you know insanely big things are going to happen, but you don't know in which direction. Then you need people who have the courage to go in directions that are uncharted territory. People who have good ideas and can look left and right and bring in new things, even if they're not experts, but have their connections. That simply worked very well with Markus. We didn't know him, he didn't know us, but that didn't matter at the time. He did his job and we drew up a plan that worked for the next four years. Uncharted Territory? A fat lot I care!
That was a nice closing. Thank you very much.
Thank you. All the best. Ciao!