It might be a long time before the general public is flying again, but the question remains: how high-risk is air travel in terms of viral infection? In an article for the Texas Advanced Computing Center (TACC), Faith Singer-Villalobos highlighted new, supercomputer-enabled research that explored how viruses travel – and transmit – on airplanes.
The new research, which was led by Ashok Srinivasan (a professor of computer science at the University of West Florida), aimed to use pedestrian dynamics models to assess disease spread in airplanes. Typically, pedestrian dynamics researchers have used the Self-Propelled Entity Dynamics model, or SPED, which essentially constitutes a molecular dynamics simulation where the molecules are people and the rules of interaction are social – not simply physical. However, like molecular dynamics models, SPED was slow, limiting its utility in urgent situations.
To bridge that gap, the research team developed CALM, loosely short for “constrained linear movements in a crowd.” CALM, which dropped the molecular dynamics framework of SPED, is targeted at assessing movement in narrow passages – and its lighter foundation allow for 90-second runtimes (a 60-fold speedup relative to SPED). The researchers applied CALM to analyze how passengers disembarked on three different airplanes. Because human behavior is unpredictable, they also ran simulations with a thousand different variables, using the distribution of the results to generate distributions of predicted human behavior.
To run their massive quantity of simulations, the researchers turned to TACC’s Frontera supercomputer, the world’s fifth largest per the latest Top500 ranking with 23.5 Linpack petaflops. Frontera’s 8,008 compute nodes are powered by Intel Xeon Platinum 8280 CPUs and connected by Mellanox HDR100 InfiniBand. Frontera also has two subsystems, both equipped with four Nvidia GPUs per node (Quadro RTX 5000s power one subsystem, while V100s power the other).
“Frontera was the natural choice, given that it was the new NSF-funded flagship machine,” Srinivasan said. “One question you have is whether you have generated a sufficient number of scenarios to cover the range of possibilities. We check this by generating histograms of quantities of interest and seeing if the histogram converges. Using Frontera, we were able to perform sufficiently large simulations that we now know what a precise answer looks like.”
The researchers made particular use of Frontera’s GPU-driven subsystem, given that CALM had been designed to leverage GPUs. “Using the GPUs turned out to be a fortunate choice because we were able to deploy these simulations in the COVID-19 emergency,” Srinivasan said. “The GPUs on Frontera are a means of generating answers fast.”
As for answers, Srinivasan cautions that models aren’t an exact proxy for real-world cases due to the impact of outlier events, but the simulations expose flaws in the systems and guide best practices.
“In our approach, we don’t aim to accurately predict the actual number of cases,” he explained. “Rather, we try to identify vulnerabilities in different policy or procedural options, such as different boarding procedures on a plane. We generate a large number of possible scenarios that could occur and examine whether one option is consistently better than the other. If it is, then it can be considered more robust. In a decision-making setting, one may wish to choose the more robust option, rather than rely on expected values from predictions.”
To read the original article discussing this research, click here.