May 16, 2019 - by Sarah Waldrip
Co-managed by ISC and the HPC-AI Advisory Council, the 2019 ISC-HPCAC Student Cluster Competition will again be an important highlight of ISC’19. The number of competitors this year has grown to an all-time high, totaling 14 teams that are traveling to Frankfurt from universities spanning multiple continents. Some teams, like ETH Zürich’s team “RACKlette,” are not traveling far, but are newcomers to the competition. Others are returning competitors with experience and previous victories to bolster their chances. All of them face the same challenge–to build and operate a compute cluster that outperforms the rest.
CSCS is a primary sponsor for the new team RACKlette, and the HPC-AI Advisory Council is directly involved with the student competition. Michele De Lorenzi, Deputy Director of CSCS Swiss National Supercomputing Centre, recently sat down with Gilad Shainer, the Chairman of the HPC-AI, to discuss in more detail the goals of the competition, what exactly the new student team can do to be prepared, and what they should expect to get out of this experience.
De Lorenzi: What is the competition about, and what is the goal of this competition?
Shainer: This is a global competition between universities to build their own supercomputer and be able to achieve the best performance with it. The only criteria, or limitation that the teams need to comply with, is 3000 Watts. Under the limitation, you can build your own innovation in hardware and create anything that you want—whether it's based on market available products or those that are not yet available in the market—doesn't really matter. Then you can innovate in the way you run the applications—both HPC workloads and deep-learning workloads—on the small supercomputer you built, while trying to achieve the best performance.
The reason we are co-managing this competition is that it helps to bring HPC and AI knowledge into universities. The student competitors practice and work hard toward a competition for nearly a year. The competition also introduces HPC into a curriculum to help students learn more about HPC. What we see years afterward is that these are the very people who find themselves in key positions, either in technology companies, or as users or hyperscalers, or within major HPC centers. And that's part of the HPC Advisory Council’s mission, to expand HPC knowledge in education.
De Lorenzi: Can you tell us more about how teams get into this competition?
Shainer: We have been co-managing this competition for multiple years now, and we also help to facilitate regional competitions months in advance of ISC. For example, we both support a regional annual competition in South Africa and host their best team in the ISC competition. There is also a competition held in China for which we reserve two spots at ISC for their winning teams. In addition, there is a regional APAC competition that we manage together with the National Supercomputing Centre (NSCC) Singapore; the top team from that competition receives a reserved spot at ISC. Beyond that, other universities can submit proposals, and we select the best from the batch. This year we are very happy to have the first team ever from Switzerland, team “RACKlette” from ETH Zürich. From competition to competition, we are seeing more countries represented over time; it's very encouraging to see how HPC is spreading around the world. We started out with eight teams many years ago, and have since grown to fourteen. We like to see teams from countries not known for having large HPC centers becoming more involved in HPC.
De Lorenzi: What makes a winning team? Is there some common characteristic amongst them?
Shainer: From my perspective, every team that participates in the ISC competition is a winning team. You need to do a lot of preparation, and you need to learn a lot of things throughout the year in order to compete. These teams not only need to build their own small supercomputing infrastructure, but also must practice running different kinds of HPC applications in advance of the competition. So once they arrive at the competition, they have already gained a lot of knowledge and brought themselves to the next level. Now, in terms of the competition itself, of course, there is some ranking, but we find that every team excels in different areas. We emphasize where each team excels at the closing ceremony of the competition, because we think it's important. We do see several teams that actually have won year after year, or usually place in the top three or top five, and I think that's because those teams come from universities that understand the value and importance of HPC. The more universities understand that HPC is technology that enables research, product development and scientific discoveries, the more often they will form the great teams and support them in the competition.
De Lorenzi: CSCS is a supporter and sponsor of the ETH Zürich team’s project. We know it’s important for students attending the competition to make contact with the other professionals, vendors and organizers at ISC, but do you have any specific words of advice for team RACKlette, who will be participating for the first time this year?
Shainer: First, thank you very much for supporting sponsorship of the team. To those young men and women, I would say the following: Some of the information about the competition’s requirements is public. Some of the workloads are known, and people can practice with as many datasets as they want. Benchmark each one of them and understand the differences or the impact of different compilers, different MPI libraries or different datasets on the performance of the application. HPCC and HPCG are known systems, so they can go and practice around them. There will be a few applications that are unknown—which they will get at the competition itself. The more they understand the impact of the compilers, the MPI libraries and the hardware, the better their capabilities to tackle the new workloads that they will not know in advance. The second thing: Always think about what could go wrong, because everything can go wrong. A server might not boot, a card might go down, something will always happen. So if you think also of what might go wrong, and you simulate those things before you come to the conference, that is always going to help. Otherwise, enjoy every minute of the competition and have some fun. Things can go wrong, but you do your best and that's the most important part.
De Lorenzi: Is there a specific expectation for the composition of each team? What are the skills you expect to see represented?
Shainer: The teams are free to define who the members are. We don't limit participation to those working on their PhDs or doing their masters or only learning their first degree. We accommodate anyone that wants to participate. We see a mixture of different people, different genders, ages, and so forth. Regarding the way that we select teams, beyond filling those places that are saved for the regional competition winners, we look first at countries that haven’t participated before. This is because we want to facilitate and bring more countries into the field, and we want to use the competitions for HPC outreach and education. We will also look at strong teams that participated in the past and are getting better from year to year. This is actually the main reason that we increased the number of teams from year to year.
- The ISC HPCAI Advisory Council Student Cluster Competition will take place from June 16-19, 2019. The Award Ceremony will take place Wednesday June 19, 2019.
- The ETH RACKlette Team will take part to the upcoming ISC Student Cluster Competition 2019 in Frankfurt.