The LUMI supercomputer is not just a very fast supercomputer, it is first of all a competence development platform - Interview with Kimmo Koski, CSC, Finland

6 Jan 2020 Helsinki - At the e-IRG Open Workshop in Helsinki, Finland,Primeur Magazinehad the opportunity to talk with Kimmo Koski about EuroHPC. Since 2004, Kimmo Koski is the Managing Director at CSC, the Finnish IT Center for Science. Before that, he gained experience at Nokia. In the nineties, he was also at CSC when he was responsible for research infrastructures. In 1989, CSC had a staff of 30 people but now this has grown to 400 people.

Primeur Magazine: We are here at the e-IRG workshop. What is your experience about the e-IRG Workshop?

Kimmo Koski:I think it is very nice. There are 60 people attending the workshop with a lot of expertise and interesting topics. It is good to have the colleagues gathered together. This gives us the opportunity to also have off-the-record discussions. During the Finnish Presidency there were organized a lot of other interesting seminars in Finland as well.

Primeur Magazine: There was a variety of topics that were addressed, from artificial intelligence to lifelong learning to EuroHPC. There were all connected by the creating competences topic. Is this also something that CSC is focusing on?

Kimmo Koski:At CSC we need of course to provide services to our users at the research centres and universities but also to people from the information management services side in research centres and the government. You cannot do that without constantly focusing on competence development. When we are growing, we need more people and we need to train them. Because CSC is a state/university-owned non-profit company, our target is to make that Finland is competitive and successful. We want to train good people. We do not mind if they transfer from us to industry as long as they pay their taxes and they employ people here.

Primeur Magazine: Are you focusing on ICT in general, or just for research?

Kimmo Koski:This is quite an interesting topic. For historical reasons, in Finland everything is centralized in CSC. Other countries have different units, hooked to universities. Research networking for instance is a different unit. In Finland, everything is under the same roof. CSC is running the national network and is taking care of the supercomputing and Cloud services, and all different kinds of other computing services, and data infrastructures. CSC is even serving the libraries, archives for museums, talking about things like open science, open data and long time preservation. More and more, CSC is also serving the government in certain areas like authentication or ICT projects. CSC has a wide spectrum of users but the origin of CSC is in high-end computing and research networks.

Primeur Magazine: You have this empty paper mill factory somewhere in the middle of the country and you thought: 'Let's put a computer there'. You are one of the three centres which will host a pre-exascale EuroHPC system.

Kimmo Koski:That is a very optimal data centre for EuroHPC. Actually, we have been running there a data centre in production since 2012. Around 2010, just for fun, we started to think about using this for computer systems. We wanted to extend our data centre outside of the Helsinki centre to somewhere else because we were using a Megawatt of computing power. The paper mill was not in use anymore, but they had a large electricity infrastructure so it turned out to be quite suitable for a data centre. Due to some interest when we were searching for possibilities, and accidental good luck, we ended up in Kajaani. It turned out to be an excellent choice.

Primeur Magazine: So for the LUMI supercomputer you brought together a consortium consisting of a large number of countries that will bring in money as will the EuroHPC JU, but also users from all over Europe. How will you handle that?

Kimmo Koski:This is a good question, of course. In our case, we have been in this business quite long. Our history started in 1971, but the first supercomputers came in the late eighties to Finland, and there is a long experience in how to get users working on those systems. Finnish science has a long tradition in High-Performance Computing and programming. But of course, it is a challenge. We have to remember that all of these three EuroHPC machines, half of the cycles belong to Europe, and the rest of the cycles depending on the funding ratio is distributed to the participating countries.

In the Spanish and Italian case, it is a different consortium, of course.

There will be quite a number of resources. We first of all want to have a maximum positive impact on science and innovation. And the other thing is we have to prepare: we want the systems to be overbooked with good codes when they become available. So we have a lot of emphasis on how to prepare the users to use the system, how they can get more out of it. And this is now one of the things we have to realise in Europe: we cannot stop investing when we have the machine. Now, we have to invest in people everywhere, not only here in Finland, but everywhere.

Primeur Magazine: From a technology perspective the vision was that researchers would have seamless access to all the machines across Europe. But with the large amounts of data involved in HPC how will that be possible?

Kimmo Koski:In some areas, the data challenge can be bigger than in some other areas, so it depends. There has been quite some discussions over the years when competing over different projects. And often the argument comes that Kajaani, where we have this data centre, located 600 km north of Helsinki, "is far" and "it is difficult with the network", but actually the fiber network is going to be there to four directions and it is actually one of the best network places in Europe. So the network is not really the problem. But of course, there are research infrastructures that are generating a lot of data, like CERN, or the SKA, but getting the data from those anywhere will be a challenge. It has to be taken case by case. What we have been doing here in Finland is that we have been running remote data centres since 2012, and that in our experience, data never has been a bottleneck. Now, with even a higher amount of resources, there will be more and more demanding challenges. In special areas we need to put a focus on data, but it will not stop it, so most of the cases it will work quite fine, like it will also work in Barcelona, or Bologna, or any of the other big centres.

At the same time we are also replacing our national systems, this year and next year, and then the following year, LUMI starts operation. So we would never have got the budget for the national systems if we were only looking for a budget in supercomputing, but we got the budget for data intensive devices with some AI flavour. So it is practically the same thing, but it demonstrates that the focus nowadays is more and more on data, not only on computing. So we will end up with some kind of challenges which we do not even know, but it will be interesting to solve them of course.

Primeur Magazine: Coming back now to the people. How will you get enough skilled people?

Kimmo Koski:As mentioned, that is one of the key things to make it work. Not only here, but everywhere. So we have this nine-country collaboration. It is about one year before the system will be in production. So with the contract stuff almost done, we need to focus on how we will together built competences. We should share, and use the synergies between the centres in co-development for example. So it is a good idea to develop a code only once, not in eight places. So there is a national need to develop competences, then there is the LUMI consortium level, and then there is the European and global level. It will be interesting to see how well we can work together with the team of these nine countries, how well can we work together with the colleagues in the Italian and Spanish centres. We are now starting to discuss about that. One thing that was clear, is that even though Finland is investing a lot in LUMI - 50 million euro is a lot on the Finnish scale - our ministry said they are not actually investing in computing resources, they are investing in a competence development platform. So what we want to do, is, of course, support the science, but at the same time make people learn new things and actually focus on this computational science and especially AI related topics too, where there is a huge need for human resources.

Primeur Magazine: In the consortium that you have, will you try to decide on special topics for each partner? So that for instance a Belgian researcher could be helped by a Polish HPC expert?

Kimmo Koski:This indeed is the idea. It is something we are now starting to discuss, because all kinds of practicalities have been dominating so far. We also start to discuss how to divide this and profile with the other centres. There is a certain overlap. When there are overlapping areas, a good idea is to try to find a way to do the things only once and share it with everybody, and then there are overlapping areas where we could complement each other. There we could bring different researchers together to invent something new. So there are a lot of opportunities, but it requires people to work and coordination. That is why we are getting all our user communities active in order to benefit from this.

Primeur Magazine: And then, of course, there has to be the coordination with EuroHPC JU and with the EuroHPC countries that are not part of your consortium.

Kimmo Koski:Exactly. With half of the machine cycles we need to support all of the European research. With our support we need to take care that all European researchers get a good service.

Primeur Magazine: On the European scale, the intention is that there is an HPC competence centre in each country, and that then these work together in supporting individual users. Is that a good model?

Kimmo Koski:It is a good model, but it is only one model. I hope it helps for part of the challenges, but there are a lot of other things we need to do to also train the users. The competence centres call just finished. And I assume most of the 30 EuroHPC countries also put their applications in. But we have not actually gone through who does what. So it will be very interesting, now the call has closed, and when the project will go public, to know what are the interests of the other countries in this competence centre call. Because this way we can also collaborate with these resources. The idea is definitely to make a kind of human network, so we can benefit from the skills of the others and vice versa.

Primeur Magazine: These centres will have a quite different expertise. There are countries with one small centre, and countries with a complex networked infrastructure of many HPC centres already. The HPC centres project will run for only two years. So it seems very complex to get it to work.

Kimmo Koski:I think it is a start. Hints from the European Commission have also been, that if it works, they will continue in the next round. So maybe we will have a continuation. But I think it is a very good initiative, but alone it is, of course, far from sufficient, because there is so much work to do. If you for instance take the maximum of 2 million euro for two years, you may get something like five to ten people, per competence centre depending on what else you do with it. We should get quite a number of more people getting involved in this. Code development alone to support acceleration of GPUs is a lot of work, for instance. But if it is done efficiently that will pay back.

Primeur Magazine: Of course, there are also Centres of Excellence in different application areas. So there is a lot of coordination to do.

Kimmo Koski:Yes, and for instance all three pre-exascale centres have to deal with GPUs, co-development and there are opportunities in AI and data analytics applications. And things are also developing, and more resources will be available in the countries. So these three centres are not the only competence in Europe. But how can we make sure that there are new engineers and researchers coming up from the education and training pipe to use and take a benefit out of these resources? So at the same time when we are focusing on HPC and high-quality research in Europe, we should start all kinds of training programmes and focus on the next generation of people who are coming through the system.

Primeur Magazine: Not only research but also users in industry could benefit from HPC. How do you see the support there?

Kimmo Koski:That is interesting. Practically, the optimal status would be that 20% maximum of the pre-exascale systems would be used for industry. Some of the centres have been successful in working with industry, other centres have done less of that. We do not have that much experience in how to work with that. This is now an opportunity to learn that and use the experience of the others. There are centres in some of the EuroHPC countries that have a lot of commercial funds. Luxembourg for instance, the Czech Republic's IT4innovation Centre is now a consortium with some industry focus. Centres in some countries are very academic. How to be able to combine these two? Think about the start-ups for example that start from academics and they end up to be industrial or commercial companies. There is an area which is actually in both sides. So we need ways to tackle those. That is one of the areas where we try to develop.

Primeur Magazine: In that sense it is a pity that neither Germany nor France, nor the UK is involved in the pre-exascale systems. Because especially Germany and the UK have a lot of experience in helping companies.

Kimmo Koski:But actually they are involved. They are not investing in pre-exascale systems at the moment, but they have their own systems. But they can of course use the pre-exascale systems, because half of the system is European. The pre-exascale centres are just one piece in the puzzle. So there will be next steps. Germany and France probably have plans for the next exascale steps of EuroHPC. But this remains to be seen. Because we are talking about the next European Framework Programme. And this is not yet decided. But this will be one of the focus areas. Germany and France are also quite active in the EuroHPC Governing Board.

Primeur Magazine: Thanks very much for this interview.

Ad Emmen