Last year, Summit was one of the first major supercomputing deployments in the fight against COVID-19, heralding a deluge that crescendoed with the cumulative efforts of virtually every research system in the world. While news around COVID-oriented supercomputing research has quieted somewhat since the introduction of highly effective vaccines, Bronson Messer – director of science at the Oak Ridge Leadership Computing Facility (OLCF) – took the virtual stage at the September meeting of the Advanced Scientific Computing Advisory Committee (ASCAC) to highlight how COVID research is alive and well on the United States’ most powerful supercomputer.
Messer opened by reflecting on the OLCF’s mission to deploy and operate the computational data resources required to tackle global challenges. “And if the COVID-19 pandemic isn’t the apotheosis of a global challenge,” he said, “I can hardly think of another.” Then, he turned his focus to another part of that mission: to work with partners to deliver transforming discoveries.
This, he said, had been embodied through the OLCF’s work with the COVID-19 High-Performance Computing Consortium, an ad hoc association of Department of Energy laboratories, major industry players, academic institutions and both federal and international agencies meant to muster the nation’s HPC resources against the pandemic.
The consortium, Messer explained, was still actively allocating resources to projects. “We still sort of have this weekly cadence, where we’re constantly reviewing proposals, trying to get matches and it is something that I can say, for me, has become part of my everyday work life every week,” he said. “We just went over the top of a hundred projects in the Consortium, and in fact, right now to be with you I am missing the current matching committee meeting.”
This continued, high pace has had serious implications for the most powerful supercomputer in the country, given Oak Ridge’s strong commitment to COVID research (several projects that were later supported by the consortium, Messer said, had actually predated the consortium through work with the OLCF). Once the consortium entered the picture, the workload was much heavier.
“If you look at our portfolio on Summit, we have 11 projects out of that 100,” Messer said. “We’ve allocated 1.42 million Summit node-hours to date on those projects. … This represents a bit more than half of all the projects that we have allocated and a bit more than half of all the allocated time … since March 2020.”
Summit’s unique strengths and ‘COVID cabinets’
Summit, he explained, continued to offer somewhat unique strengths in pandemic research. “Summit is the most powerful hybrid CPU-GPU machine ever built,” Messer said. “There are a handful of things that it’s really really good at. Importantly, it can do anything.” But with some applications, he said, the machine’s strengths – predominantly “an inherent, almost overwhelming advantage” in molecular dynamics – put it above the rest of the pack from the get-go.
“Then,” he said, “something interesting happened. In the middle of last calendar year – end of June, very beginning of July – we installed a handful of new nodes in Summit. We expanded Summit with CARES Act funding to provide 54 more nodes.” The nodes “are a little bit different than the rest of Summit,” he said, explaining that they each have double the high-bandwidth memory on the GPUs and four times the DDR4 and non-volatile memory.
“Because of that, we actually run them as a separate queue, a separate partition of the machine,” he continued. “We can put them back in the general population if we really, really want to, but it turns out that running them as a separate partition has been very advantageous for a handful of projects because we’re able to offer this special queue with these special characteristics versus the rest of the machine – especially projects that, for example, need to train on very large datasets but are not data-parallel … These nodes are perfect for that kind of work.”
Research from Jennifer Diaz (now of UCLA), he said, exemplified the benefits provided by these nodes. Her research worked to predict synergistic drug combinations for COVID-19 treatments using a pilot set of 22,000 drugs. Now, Messer said, “she’s planning a one million drug screen in the near future because we were able to enable this workflow effectively on these ‘COVID cabinets’ on Summit.”
Looking ahead to more organized urgent computing
But working with the consortium revealed more than scientific insights – it revealed how quickly diverse, urgent science could happen in a collaborative HPC environment. “From my perspective as the person who is responsible for making sure that the allocations on the machine are distributed according to various priorities,” Messer said, “I think that the most interesting part of the consortium was the ability of us at OLCF to ‘open our aperture’ to projects that perhaps wouldn’t have made their way to us beforehand, and that we could uniquely impact on the way to getting scientific insight.”
Now, of course, the HPC world is buzzing with debate over the possibility of a National Strategic Computing Reserve (NSCR) – which would be like the COVID-19 HPC Consortium, but prepared in advance for a variety of urgent computing use cases.
“I think that the lessons that we’ve learned during the HPC Consortium go a long way to inform us,” Messer said, speaking to the possibility of an NSCR. “We now have an existence proof, so I think we’re content with the fact that we could do such a thing.” But, he hedged, there were big, open questions: “What does a reserve mean? It cannot possibly mean that we’re actually holding back time – it has to mean that we’re able to quickly deploy new allocations, stand up new projects, get them enabled, get them running as fast as possible. I think that the coordination that we’ve already shown through the consortium says that this is something that actually could be done – but it will require an understanding of all the possible providers’ needs, their policies [and] ultimately, their missions.”