On 22 January 2019, we were invited at the EuroHPC Joint Undertaking (JU) premises in Luxembourg to talk with Gustav Kalbe, head of the High Performance Computing and Quantum Technologies Unit at the European Commission, who currently is EuroHPC JU's interim Executive Director.
Can you explain your role in the EuroHPC Joint Undertaking?
Gustav Kalbe: The EuroHPC JU was legally established in October 2018 and is now in the process of establishing the working and operations of the Joint Undertaking which includes the appointment of the Executive Director. Until that post is formally occupied through an external selection process, I am the interim Executive Director responsible for establishing the Joint Undertaking, setting up the office, hiring part of the staff, establishing all administrative and legal procedures, implementing the activities and decisions that were taken by the Governing Board. Hopefully by the end of the year with the nomination of the final Executive Director, the JU would become autonomous. I will no longer have the double role of on the one hand being the Executive Director and on the other hand working at the European Commission as head of the unit that is responsible for the HPC policy in the Commission.
On 21 January 2019, the first EuroHPC Call has opened, announcing the first real activities of the Joint Undertaking.
Gustav Kalbe: Clearly, we spent the first three months in setting up the key processes to be operational and be in a state to publish the first Call of the Joint Undertaking which is the first tangible action to be implemented by the JU. This happened on 21 January 2019: the Call for Expression of Interest for the Selection of Hosting Entities for Precursors to Exascale Supercomputers. This is the first step for acquiring and putting in place the pre-exascale supercomputers that will be owned by the JU and made available to the user community at large in the Union.
Building an HPC ecosystem
What are the main objectives of the EuroHPC JU?
Gustav Kalbe: Let us separate them in two. The first objective is to install and operate a world class supercomputing infrastructure which will be in part be built around the pre-exascale supercomputers that the JU will acquire, complemented by a number of petascale machines. Those machines will be interconnected together but also, where possible, connected to national machines and even PRACE. This objective of making available world class machines and sufficient computing time is complemented by a research programme to prepare the users to develop the key technologies, the key software and applications to continue staying at the forefront of computing performance and having the next generation of machines acquired around 2022-2023 with the exascale area and then moving to the post-exascale area. This is the second objective. We are not just running a one-off exercise but have long-term objectives. We are constantly planning to renew and expand the computing capacities and always, in parallel, prepare the grounds for the upcoming generation where on the one hand we develop the required technology but also are ready to exploit the performance of the machines.
The applications that will be in the research programmes, are these the same kind of applications that we are running today or are there other application areas involved?
Gustav Kalbe: This is an excellent question which we are addressing through different angles. There are a number of applications that definitely need the exascale. These are the applications we want to develop and did not exist. We also want to identify the ones that are upcoming, as well as determine which of the current applications need to be ported or make sense to be ported to the exascale. We know that there is a number of applications that, for the moment, do not really make the case to move to exascale. Therefore, we now just have launched as a result of the last Calls for proposals of Horizon 2020 a new round for Centres of Excellence that are focusing on a number of key application areas where there is a clear justification to move towards the exascale. From all that we know, and experts agree, it is not just taking existing codes and moving them to next generation machines. There is a whole set of new developments that need to be done. The Centres of Excellence are needed to bring the various user communities around these applications together to prepare the codes and models for the applications that can be run on the exascale later on.
This is not only for scientists but also for companies?
Gustav Kalbe: The JU is using public money to buy an infrastructure. Therefore, primarily, its usage should be public research. Irrespective whether it is for a university, a research centre or an industry, as long as it is public research, anyone can use it free of charge. However, the EuroHPC regulation foresees the possibility to use the time of the machines up to 20 percent for commercial applications, mainly from industry, and hopefully for a large proportion from SMEs, without going against the Framework on State Aid on private or commercial usage of public infrastructures. If you use the machines also for commercial purposes, those users would have to pay market prices to get access to the machines to avoid distortion of the market. This is clearly established in the EuroHPC regulation. As for the last set of users, we are seeing more and more interest of bodies and administrations, ministries and so on, to also use HPC for their purposes. This is also a planned constituency that we would like to expand as user base.
This is not like a one-time moonshot, it is not like "We want to do exascale and that is it". There is a complete programme and a complete ecosystem?
Gustav Kalbe: Indeed. The aim is not to build a champion to put in a vitrine in the museum to show we are capable of building the fastest machine in the world. There has been a consensus from the very beginning that we are not focusing on machines but on building a whole ecosystem which on the one hand builds on making available across Europe, irrespective of the location, access to world class computing facilities in an interconnected way. The idea is that the user, wherever he is located in Europe, will have a one-stop shop for his needs. He will go to this centre-of-contact point, a local contact point, with the request: "want to run this code". Depending on the needs or the demands of the application, it would be distributed on a free machine in the EuroHPC ecosystem. This could be on the fastest machine because it is a very demanding application or if the application does not demand that kind of power, it can be located somewhere else on the available machines. For the user, it is completely transparent. He does not need to know where it will happen. Precisely, by interconnecting the machines, we will create a complete offering of necessary capacity depending on the applications, depending on the performance needs in a way that we do not need to send people around, saying: "For this programme, you have to go to that place and talk to the experts there". Don't forget that all the machines that will be acquired by the JU are placed under shared ownership. For the pre-exascale machines, half of the computer time would be made available for the users across the Union for free, the other half would be for the national use for the countries that have invested collectively in that machine. For the petascale machines, up to 35 percent of the capacity would be made available to the Joint Undertaking. There are some countries that in addition might voluntarily offer 20 percent of the national machines so that in the end the JU would somehow manage access to a number of machines, spread across Europe, which of course have to be connected. In this way, we can make optimal usage of the capacity.
The other part of the ecosystem is more the development side. In order to stay at the forefront of having machines that optimally use applications and really drive most out of it that is world leading in application performance, you need to have the co-design of the hardware, the users, the supercomputing centres, the software, etc. You have to bring all actors together in this ecosystem to make sure that at the end you have machines that are achieving most of their performance because they have been built like that.
And how does it work for the petascale machines?
Gustav Kalbe: The core idea behind the petascale systems is to complement the infrastructure with additional capacity. This will also help the Participating States which have already certain competences in supercomputing to raise their levels to come closer to the pre-exascale. As for the pre-exascale however, we have the possibility to have a consortium. Several countries can say: "We all want to contribute to the funding of the machine be it the acquisition or the operating costs; we create a consortium and the machine will be located in one country but we have a shared ownership". Let us take an example of three countries coming together. Each of them would for example have one third of the machine for their own national usage. The other part which is funded by the JU is for free usage for any user in the Union, irrespective of the fact that his country made a contribution to the acquisition or operating costs of the machine and irrespective of the fact that the country may or may not be part of the JU. The third part of the ecosystem is more the development side. In order to stay at the forefront of having machines that optimally use applications and really drive most out of it that is world leading in application performance, you need to have the co-design of the hardware, the users, the supercomputing centres, the software, etc. You have to bring all actors together in this ecosystem to make sure that at the end you have machines that are achieving most of their performance because they have been built like that.
Centres of Excellence and Competence Centres
Can you tell a bit more about how the Centres of Excellence are going to help?
Gustav Kalbe: In the first round, we had seven HPC Centres of Excellence. They are coming to an end now. Recently however, nine new ones are up and running.
They are supporting applications areas?
Gustav Kalbe: They are supporting user communities. Their mission is twofold: to bring the users together so that they can benefit from the applications being developed, but secondly, which is more important, to prepare the user communities for the exascale area to develop with them the codes, the software that we need to run the exascale machines and take most out of the performance offered by those machines.
When you say that people should go to their local centre to be connected to the correct machine, then you are talking about another type of centre?
Gustav Kalbe: Yes, this is a different story. These are the Competence Centres. We would like to see at least one in each country. This would be the local presence helping the user communities to understand what HPC is all about, what it can bring for their particular problem, helping them identifying the solutions, providing them access to the ecosystem, the machines, the applications, and so on, and then following them further through the whole project on exploiting the HPC capabilities.
You want to support users on the road to exascale, in developing HPC applications and helping them when they are actually doing practical work. Then there is also the hardware infrastructure that will be acquired in part by the JU, together with some of the countries. This is complemented by the technical development of the new hardware. What are the plans there?
Gustav Kalbe: If you read the various communications and you look into the EuroHPC regulation, the key objective is to create the exascale ecosystem. In order to come there, we need of course to prepare the technology. We also said the ambition is to have at least one of the exascale machines built for the majority out of European components to maximize the outcome of the research portion of the funding. In parallel with the launch of the Call for selecting the hosting entities where the pre-exascale machines will be located, we are planning to launch the technology development so that in 2022, when we will organize the Call for the procurement of the exascale machines, we will have as much as possible indigenous technology that can be incorporated in the machines we are going to buy.
Is it still in the planning to have two exascale machines?
Gustav Kalbe: That is what we said in the objectives: to have at least two machines. Now, we cannot be very concrete on the final number of exascale machines because for the moment the Regulation of the JU does not foresee the budget for the exascale machines. We have to wait for the next financial Framework of the Union, to make an amendment to the Regulation and see, depending on the budget that we will get the complementary funding from the participating Member States, before we can see how many exascale machines we could buy. Whether that will be two or three or four, this will depend on the price of the machines and on the available budget, but the ambition is to have at least two exascale machines.
Tentative EuroHPC Calls
And preferably one with as much indigenous technology as possible.
Gustav Kalbe: Yes.
For the development cycle, there will be also new Calls coming out. Can you say something about that?
Gustav Kalbe: The strategy for the HPC initiative dates back already a couple of years. The first concrete attempt to have a structured initiative was at the beginning of Horizon 2020. It is now the role of the JU to take up the different steps and continue the development roadmap. That is why the JU will publish a number of research Calls that are building on the previous Horizon 2020 Calls to implement a long-term roadmap towards the exascale. We have planned to have a number of Calls in 2019. The next ones are in 2020 to drive this roadmap to build on the previous phases. For the moment, we are still having to decide with the Governing Board what will be the scope of the objectives in 2019 and 2020, what will be the respective budgets, knowing that there is a certain flexibility, although limited to a certain extent because we have this continuity aspect that certain Calls are a "must have". We have to continue some activities if we don't want to miss the exascale objective or to lose the previous investment. The JU is now counting on to get from the Member States additional contributions to double this budget. The Workprogramme can now be much more ambitious and add other topics that we have not foreseen in Horizon 2020.
But of course, time is ticking away just as fast as it did before. That is why you cannot wait too long.
Gustav Kalbe: No, the idea is that we are having the Call out for the selection of the hosting entities and now we are working on the research Calls. Before the end of the year, we have to publish a number of them, hoping we can do the first one before the Summer.
Do you have an idea of the space between the time of publishing and between the time of a Call closing?
Gustav Kalbe: There we are bound in following the Horizon 2020 rules. The minimum is two months between Call opening and Call closure.
There will be a bigger budget because it will be matched with national funding but this also comes with a price. When you submit a project proposal it will be evaluated in two places? How will this process go?
Gustav Kalbe: Trying to make it simple, the JU will cover up to 50 percent of the costs of the projects that are retained. The other half has to be provided, either by the Participating States of the JU that have committed budget to that specific Call for their national beneficiaries, or in the absence of national funding, will have to be found from other resources. This complicates the life of the potential beneficiaries because they would have at least two funding sources, one following European Horizon 2020 rules, and one following the national programme which differs from country to country in terms of funding rates, eligibility criteria, and whatsoever.
In the worst case it could be like twenty partners from twenty different countries with twenty-one different rules applying.
Gustav Kalbe: Indeed, and some of them might not even get any national funding because their country has not set aside a dedicated budget for this specific Call.
Is it known on beforehand?
Gustav Kalbe: Yes. Otherwise it would not work. The core idea - and this is inspired very heavily on how the ECSEL JU is operating - is that at one stage, the JU is going to publish a Call. We say: these are the topics that are now open; these are the objectives; this is the budget that is committed from the JU, to cover 50 percent of the costs; here are the other countries that are committing "xyz" amount, each of them for the different Calls.
If a country doesn't commit, a beneficiary could still participate but then he has to pay himself.
Gustav Kalbe: Here, the Horizon 2020 rules for participation apply, irrespective of the fact whether a country is or is not a member of the JU and irrespective of the fact whether a country is contributing or not to the budget. The only difference for the beneficiary will be that the absence of a national commitment for the selected Call will mean that he will only get 50 percent funding from the JU. The other 50 percent, he will have to find from other sources.
That sounds simple again.
Gustav Kalbe: Yes, it sounds simple but in practice, it is not so straightforward.
Can you summarize in a few sentences the working and operations of the JU?
Gustav Kalbe: Let me work backwards. For the next two years, what we plan to do, is by the end of 2020, to have the pre-exascale and petascale machines installed and operational. In order to do so, we have to organize the acquisition phase which we can only start, once we have decided where the machines that will be bought with the JU, will be located. That is why we launched on 21 January 2019 the first step of this infrastructure pillar of the JU by publishing the Call for Expression of Interest for potential hosting entities to make a bid if they want to host and run the machines on behalf of the JU. In parallel, we plan to launch this year a number of Calls for the research activities of the JU to prepare what comes after the pre-exascale machines, namely the exascale machines which we can hopefully procure in 2022-2023.
That is a nice summary. Can you tell something about how the organisation of the JU will develop?
Gustav Kalbe: The JU is now in a transition phase. We are currently relying on the resources of the European Commission that initiated the whole initiative and is responsible for the policy in that area. The first step was for the Commission to appoint an interim Executive Director which, traditionally for the JU, is the Head of the Unit that is responsible for the file. For the time being, I have a whole team that is working full time and very hard to get everything in place and implement the actions of the JU. We have now the possibility to launch the recruitment process of the staff, including the Executive Director. We are negotiating with the country of Luxembourg the seat of the office. We are also negotiating the various contracts and service level agreements needed to be autonomous by the end of the year.
What is the amount of the full time employees (FTEs) we are talking about?
Gustav Kalbe: For the next three years, we are talking about a maximum of 16 FTEs. This is specified by the financial fiche that accompanied the proposal to the Commission. That is why I hope that pretty soon we can start that recruitment process which will be an external process. For the moment, I cannot say anything more about that.
(Disclaimer: The views expressed in the article are the sole responsibility of the author and in no way represent the view of the European Commission and its services.)