CSCS Hosts 74th HPC User Forum in Lugano

“Switzerland is fully committed to making positive contributions to supercomputing now and in the future,” CSCS Director Thomas Schulthess said in the opening moments of the forum, before introducing Peter Brönnimann, State Secretariat for Education, Research and Innovation, to deliver an update on the state of HPC in Switzerland. Brönnimann discussed why Switzerland was joining efforts with other nations in EuroHPC and outlined goals for achieving exascale by 2021 – 2024.

Schulthess also gave a presentation that focused on CSCS’s involvement moving forward with not only EuroHPC, but also the “pre-exascale ecosystem” LUMI at a datacenter in Kajaani, Finland. EuroHPC and the LUMI Consortium plan to build the first Pan-European Supercomputer in this location. Later in the day, Kimmo Koski, Managing Director of CSC, a non-profit center of expertise in information technology owned by the Finnish state, delivered an extended presentation on the LUMI project and described its many benefits—not just to the European HPC community, but also to the local inhabitants of Kajaani, as the excess heat from the machine can be used to heat the city. The computer is projected to be the most powerful supercomputer in Europe upon completion, and it will effectively have a negative CO2 footprint.

Simon Burbidge, Director of the Advanced Computing Research Centre (ACRC) at University of Bristol, UK, gave the final presentation of Day 1, offering attendees a closer look at ARM’s Marvell ThunderX2—a competitive new 64-bit CPU in the HPC market. “This is brand new CPU technology, never been used in HPC before,” he said in the presentation, “and it works.”

Smaller scale ARM processor designs have been widely licensed to private companies for use in common electronics—cars, washing machines, smart watches and phones, for example—because they are, by design, easy to customize for specific applications. However, the “Isambard” supercomputer, the first ARM-based supercomputer (Cray XC50 system), was launched in 2018 in Bristol, and it has since proven that the processors are performance competitive for mainstream HPC workloads. Furthermore, Burbidge noted there have been positive results with the UK Catalyst Programme as well, which is a collaboration between Hewlett Packard Enterprise (HPE), ARM, and three UK Universities. Both Catalyst and Isambard are fully functional ARM-based systems doing real user workloads since January 2019.

Since making a splash at ISC 2019 in Frankfurt over the summer, the UK-based team has continued to port more application codes to the system and is now hosting approximately 1,200 users and 500 projects. With some benchmarks met and real science now being carried out on the system, Burbidge explained that it was time to start reaching out to a wider market. “The HPC User Forum has a larger industrial and commercial membership than most of the other conferences, which tend to be mostly academic,” he says. “So I wanted to get the message out there for them that there is potentially a new kid on the HPC processor block.”

“CSCS has become one of the world’s leading supercomputer sites, and one of the key reasons for that is that even though CSCS hosts a lot of science in different areas, weather is one of the most important ones,” Conway says. This is why the Steering Committee felt that Lugano would be a good location for a forum that provides extra focus on weather and climate science in HPC. “The work here for the Swiss National Weather Service is very interesting to people in the weather community, because they moved it onto a very new kind of technology—the GPU technology—and they are having very good results with it.” Conway also notes that Switzerland’s weather is notoriously difficult to forecast due to the extreme mountainous terrain, making it an enticing place to test weather prediction technology. “The whole global weather community has been eagerly watching what’s happening here.”

Rob Aulwes, visiting from Los Alamos National Laboratory, USA, added to that excitement with his presentation on possible new strategies for GPU-based climate modelling. Specifically, he stressed the importance of “concurrency”—using both GPU and CPU processors simultaneously to get the most performance from each computing node. He explained that the team at Los Alamos has achieved significant progress, but it’s still a matter of trial-and-error experimentation. New areas must be identified for acceleration. One possible way forward, for instance, is dividing the workload for a global climate model so that GPUs would be used for ocean modelling (on 6 cores plus 6 GPUs) and CPUs would calculate sea ice movement (on remaining 34 cores per node).

Trending Topics: Artificial Intelligence and Cloud Computing

The uses of artificial intelligence (AI) are growing exponentially in today’s world. From medical devices to search engines—when it comes to research studies on a global basis, there’s a pretty strong consensus: “At the forefront of developments in artificial intelligence, HPC is indispensable,” says Conway. He describes the situation as a convergence of interests from the scientific community and large commercial companies, in part because less complex forms of AI are in increasingly high demand for every-day devices in the private sector (like Amazon’s Alexa, for example). These companies, like Amazon, Google, and others, are now an interested party in discussions going on at places like the HPC User Forum.

“It’s good for them, and it’s good for people interested in HPC,” Conway says. “If they want to be leaders in artificial intelligence, which is crucial to their business, then this is the market that will get there first. This market is going to lead the way.” Every single automobile maker uses HPC to design cars, and every weather service uses it to make forecasts. The market has grown from 2 billion dollars in the 1980s to about 30 billion dollars today, Conway notes. “This is an extremely high-growth market, plus a time of quick change with the advance of artificial intelligence and cloud computing.”