The Texas Advanced Computing Center (TACC) and its juggernaut Frontera system (ranked ninth on the most recent Top500 list) have been on the front lines of COVID-19 research for around eight months – and Frontera itself had only entered full production six months prior. Dan Stanzione, associate vice president for research at the University of Texas at Austin and executive director of TACC, has been there for all of it.
In advance of some big news for Frontera (more on that later), Stanzione sat down for a digital “fireside chat” with Nash Palaniswamy (general manager for AI and HPC solutions and sales at Intel) to talk about the ways TACC has adapted to remote work, COVID-19 research and changes in computing over the course of 2020.
“I think we’ve had a really outstanding first year [with Frontera],” Stanzione said. “It’s been full the whole time. Tremendous demand, very high utilization, great reliability.” Frontera, he said, had supported over a million jobs across “70 to 80” teams, ranging from single-node jobs to jobs that used around 7,900 of the system’s 8,008 Intel Cascade Lake-based nodes. These jobs have ranged from tornado dynamics to hypersonic aircraft to materials design – and, Stanzione said, “this year in particular, an awful lot of drug discovery.”
“We’ve had a ton of use related to COVID,” he continued. “In fact, beginning maybe the last week of February or so and running through maybe the end of August, about 30 percent of what we were running was COVID-related one way or another. It started to drop off a little bit as we made it through the early stages of some of the projects. Across TACC, we’ve supported 50 to 60 COVID-related projects – 28 of those have been on Frontera this year.”
“… about 30 percent of what we were running was COVID-related one way or another.”
The projects, Stanzione said, had included everything from the microscale (“we do whole-virion modeling, we do a lot of molecular docking, look at the atom-by-atom level of how the virus moves, how it gets into a cell, how you might bind to it in order to build vaccines”) to the macroscale (“looking at society and people where we’re looking at mobility data, cell phone-related data, mapping epidemiological models to how much people are moving around and interacting”). In fact, he said, TACC still uses Frontera to run daily forecasts of those epidemiological models for policy-makers.
By way of example, Stanzione highlighted work by Rommie Amaro’s team at the University of California, San Diego.
“The first thing that we did was characterize the spike on top of the virus particle – that’s what everyone was doing – and then you scale that out to a few hundred million atoms and you get the whole virion,” he said. “And then you look at that over time, because that spike actually wiggles around some, and that motion turns out to be really important. [Amaro] discovered that that spike kind of hides in a coating of sugar.”
Understanding that the spike emerged from this coating “every four microseconds or so,” he explained, was key to designing effective drugs. That work has since been utilized by researchers like Rick Stevens at Argonne National Laboratory who are working to use hybrid simulation-AI pipelines to engage in rapid drug discovery. “This blend of simulation and AI has really proven to be very powerful and perhaps more effective than either one of those techniques would have done on its own, particularly in the drug discovery case,” he said. (To learn more about the COVID-19 research from Amaro and Stevens, read HPCwire‘s coverage of their SC20 panel.)
“This blend of simulation and AI has really proven to be very powerful and perhaps more effective than either one of those techniques would have done on its own.”
“I think we’ve put easily ten million node hours of Frontera time into COVID-related work at this point,” Stanzione estimated, adding that in addition to 30 percent of TACC’s computing time, COVID-19 research had occupied around a third of TACC’s staff. The COVID-19 research has also been making full use of Frontera’s various technologies – its Intel Optane DC persistent memory, for instance, was used to effectively create “multiple terabytes of RAM” for a COVID-19 project conducted by the Cleveland Clinic.
Frontera, of course, is not the only system at TACC. Second most-notable is the Stampede2 supercomputer, a 10.7-Linpack petaflops system that debuted just two years before Frontera. Stanzione said that while Frontera has taken over the “few largest computational problems,” Stampede2 is even busier: it recently crossed the 2,000-project line, he said, and in August, hosted its 8,000th user. Stanzione assured that Stampede2 – which has hosted over five million jobs across four years – has “two good years” ahead of it before TACC starts thinking about Stampede3.
Stanzione also commented on how remote work and the pandemic generally have changed the workflow at TACC.
“We’ve dealt with what everyone has dealt with in terms of the pandemic and switching to largely remote work,” he said. “Obviously we’ve kept a footprint here in the building near the datacenter all the time, just to keep an eye on the hardware. … But most everybody else has transitioned to remote work. Given the nature of the computational work we do, I think we were well-prepared to do that.”
“While we’re seeing the huge volume of demand, we’re also seeing different kinds of demand,” he continued. TACC, he explained, had been spending a lot of time building out web and other interfaces for its systems to smooth out remote workflows. “We have a project now where we’re using Google Docs and natural language processing as a way to specify computation on the machine.”
Accordingly, TACC has been hiring in a variety of areas during the pandemic, with its total staff now numbering over 180. “We’re always hiring great people,” he said, “so for anybody out there who’s looking for something to do, TACC might have a place for you.”
And TACC’s staff isn’t the only thing getting upsized: Frontera itself will also be receiving a major upgrade, announced this week at SC20. The upgrade will add 396 Dell R640 server nodes, each containing two Intel Xeon 8280 Cascade Lake CPUs and 192GB of DDR4 memory: an identical configuration to Frontera’s existing 8,008 nodes, which deliver, in aggregate, 23.5 Linpack petaflops. The new hardware provides an additional ~1.15 Linpack petaflops, although TACC currently has no plans to redo the benchmarking. The expansion is aimed at increasing Frontera’s capacity for urgent computing like COVID-19 research and natural disaster analysis.