Experts from the HZDR help to launch the “Frontier” supercomputer in the USA
The world’s fastest supercomputer is currently being built in the US state of Tennessee: the companies Cray and AMD are installing it at the Oak Ridge National Laboratory on behalf of the US Department of Energy until 2021. Thanks to an innovative graphics processor architecture, “Frontier” will be able to solve more than one and a half trillion floating-point computing tasks per second. It is probably the first computer to enter the Exaflops class. Physicists from the Helmholtz Center Dresden-Rossendorf (HZDR) will be among the first users. They form one of eight selected international teams. In cooperation with the project leader Prof. Sunita Chandrasekaran from the University of Delaware, the Dresden scientists want to develop scientific pilot tasks and make the novel supercomputer more user-friendly for researchers from all over the world.
The American-Saxon cooperation is based on the success the experts from Dresden have gained internationally through their experience in particle simulations and supercomputer programming. “Frontier will break a sound barrier,” estimates Dr. Michael Bussmann, head of the “CASUS – Center for Advanced Systems Understanding” department at the HZDR. “We can be proud that our colleagues from Oak Ridge have invited us to accompany them on the journey into a new scientific and technological territory,” emphasizes Dr. Guido Juckeland, who is the head of Department of Computational Science at the HZDR and, like Michael Bussmann, part of the Frontier team.
Oak Ridge expressly welcomes suggestions for improvement
In order to make their novel supercomputer quickly usable for science, the Americans have established a “Center for Accelerated Application Readiness” (CAAR). The Oak Ridge Leadership Computing Facility (OLCF) of the U.S. Department of Energy has now invited eight groups of experts from around the world to help in the start-up phase of Frontier. Each group will run simulations that can only be run on an exaflops class supercomputer. At the same time, each simulation will help to solve a particularly challenging scientific problem. One of these teams is the University of Delaware and the HZDR.
The US colleagues have also requested these international collaborations because their frontier has a few special features. These include its digital components: for the first time, graphics processors from the US company AMD are being used for a high-performance computer of this size. They are considered very powerful in the world of normal PCs. However, there is no experience worldwide of building exaflops supercomputers from these special chips. The Dresden experts will help to get a grip on the expected initial problems.
Overcoming boundaries between Intel, AMD and Nvidia
The HZDR research group around Michael Bussmann has developed special expertise in scientific software over the years. With their customized programs the researchers can simulate the interaction of ions and other tiny particles on neutron stars or in superlaser experiments particularly efficiently – and they do so on supercomputers with very different designs. Their software packages “PIConGPU” (Particle Simulation in Cells on Graphic Processors) and “Alpaka” are regarded as pioneering in this respect. “Thanks to our codes, such simulations run very efficiently on very different hardware platforms,” Bussmann estimates. The HZDR researchers have already adapted their program libraries to high-performance computers that use Intel, AMD or ARM main processors or are built from Nvidia graphics processors. For Frontier, they are now optimizing their software for supercomputers made from AMD graphics chips – this is new technological territory.
The HZDR simulation software “PIConGPU” is designed to answer current questions in accelerator physics. Dr. Alexander Debus and Dr. Thomas Kluge from the Institute of Radiation Physics at the HZDR, for example, are working on innovative concepts for high-intensity lasers with which light electrons and heavy ions can be accelerated far more efficiently and space-savingly than is possible in today’s linear and ring accelerators. The laser-driven plasma accelerators can achieve the maximum electron energy of linear accelerators kilometers in length on a laboratory scale.
Compact accelerators against cancer
“We think that we will be able to achieve beam energies beyond ten giga-electronvolts in a single pass without having to reset the electron accelerator several times,” explains Debus. “In simulations we want to show that we can overcome the old limitations. But this will require very powerful computers like Frontier.” Debus and Kluge want to use supercomputers to investigate the complex physical phenomena during such a long accelerator run. The first prototype of the new laser accelerators will also first be built in the virtual supercomputer world before construction begins in the physical world. Conceivable fields of application for such laser-driven ion and electron accelerators are, for example, the treatment of cancer using proton therapy, particle research or astrophysics.