Developing own EuroHPC hardware could make all the difference for exascale supercomputing in Europe

19 Apr 2018 Sofia - Thomas Lippert from the Juelich Supercomputing Centre chaired a session about HPC applications and their impact on Europe at the EuroHPC Conference in Sofia, Bulgaria on April 19th, 2018. He started with an overview of PRACE, EuroHPC and the importance of supercomputing in Europe.

First he took a look into the past. What has supercomputing actually achieved? We have achieved ground breaking results in science in Europe with supercomputing, Thomas Lippert said. PRACE delivers supercomputing science across Europe through a peer review process. There are 5 hosting countries of PRACE supercomputers. These will be boosted to (pre)exascale through the EuroHPC programme. This is mandatory for the success of European science and industry.

The European Union has supported the development of the PRACE infrastructure through a series of six pan-European implementation projects (PRACE-IP projects). Currently there are 26 countries member of PRACE.

The agenda of Thomas Lippert's presentation looked as follows:

  • Introduction into PRACE: Europe's supercomputing partnership is unique in Europe.
  • How PRACE enables leading edge simulation projects in Europe
  • PRACE as an element in the European strategy for the exascale era.

PRACE has a long history of over a decade now. It provides top level (Tier-0 ) supercomputing capacity collected from amongst the fastest national supercomputers in Europe. The next level are the national (Tier-1) national supercomputers, and at the base of the computing pyramid are the Tier-2 regional supercomputers. So there is a complete ecosystem in place.

Apart from the PRACE projects that each are tens of millions of euros in size, there is a small PRACE AISBL organisation supporting the PRACE developments.

There are now 7 supercomputers in 5 countries that are part of PRACE. One of these machines is in the TOP10 worldwide.

"But to become really excellent, we need to ground this by very big systems. That is the reason why we want to go for EuroHPC", Thomas Lippert stated. "In the future, EuroHPC will provide pre-exascale and exascale machines."

Until now PRACE has supported over 570 scientific projects, and provided some 16 billion core hours since 2010. Of course, a 2010 core is not comparable to a 2018 core. PRACE also supports SMEs through its SHAPE programme.

Some examples of leading edge science developed thanks to PRACE are:

  • Modelling gravitational wave signals on the Mare Nostrum supercomputer. Served as templates for the analysis of the LIGO/Virgo detectors.
  • Carbon nanotubes as excitonic insulators on the Marconi supercomputer. Proved from ab-inition calculations.
  • Convection resolving climate simulations at the Piz Daint supercomputer. It runs simulations 2-3 times faster that the TaihuLight system, the current world #1.
  • Grand challenge of drug design on the Piz Daint reduced production run costs for one drug to 10.000 euro on the Piz Daint supercomputer. A reduction of a factor of 1000.
  • Biggest crash optimization ever on the Curie supercomputer. It models 20 million degree of freedom finite element models.
  • HPC welding on the Hazel Hen supercomputer. An example of an SME application about multi-layered welding processes.
  • Basic science studying the difference between the mass of the neutron and proton by using QCD simulations on the JUQUEEN supercomputer.
  • Supernova simulation on the SuperMUC supercomputer. It predicts a new neutrino-emmission assymetry in forming neutron stars.

PRACE is a centerpiece to support Europe's Digital Strategy in the exascale era, Thomas Lippert continued.

We have a top infrastructure, EuroHPC, which is a huge investment, and PRACE which is a community effort about top science. The EuroHPC infrastructure is managed top-down. PRACE is more a bottom-up development. EuroHPC focuses on acquisition and operation, PRACE focuses on provisioning. EuroHPC looks at research and innovation, PRACE at training and user support. So, the two are complementary. Thomas Lippert sees it as a goal to create optimal synergy and complementarity.

Thomas Lippert expects a full pre-exascale system will be in operation by the end of 2021. It is not utopian to say an exascale computer could be operating by the end of 2023. But it is not just one date; it is a process with several phases.

Answering a question fromPrimeur Magazineabout whether 1 billion of investments in EuroHPC exascale systems will be enough, and whether it would be better to just buy hardware and concentrate on applications, Thomas Lippert said we will do all of that in Europe. There will not only be European hardware, but the major point is we also have the chance to develop better hardware in Europe. If you want to get onto this level of competence and you buy this type of supercomputing hardware in China or the US, you pay the same amount of money as you pay for developing it on your own. So it does not mean so much difference in spending, but it creates so much difference in creative know how.

Ad Emmen