The coronavirus pandemic turned Folding@Home into an exaFLOP supercomputer

Almost 20 years ago, faculty in the chemistry department of Stanford University launched a distributed computing project called Folding@Home (F@H). They sought to understand how proteins self-organize and why this process sometimes goes wrong, causing issues such as cancer and Alzheimer’s Disease.

F@H hit its pinnacle of mindshare—and performance—in 2007, when Sony added it to the PlayStation 3. But like many other projects, it saw a gradual decline in its popularity since. This past March, however, F@H saw a sudden resurgence. Thanks to a confluence of events, notably including the SARS-CoV-2 pandemic, Folding@Home broke the exaFLOP barrier at least one or two years before Intel, AMD, IBM, or Cray could do it. Here’s how those events played out.

In distribution

Distributed computing projects partition a massive processing job out to individual computers, with each doing a small slice of the job. You can set the most distributed computing apps to run only when your PC is idle or let it run in the background while using it.

SETI@Home, which was one of the first high-profile projects, sifted through radio telescope recordings to look for extraterrestrial signals. GIMPS, or the Great Internet Mersenne Prime Search, sought to find the largest prime numbers ever. Distributed.net tried to break RSA encryption. And so on.

The higher-profile SETI@Home recently ended its 21-year run having found nothing. By contrast, Folding@Home yielded quite a few results, in the form of 233 research papers. Among the big finds were studies into the dynamic nature of kinases and G protein-coupled receptors (GPCRs), as well as work on antibiotic-resistant bacteria and proteins from the Ebola virus.

F@H learned new ways to target these proteins in structure-based drug design, and findings from the simulations are being used by labs and startups, according to Dr. Vijay Pande, the former Stanford chemistry professor who led the project for 19 years. Today, Pande is a venture capitalist with Andreesen Horowitz. “One of the reasons I changed [careers] to investing is because we have shifted from the basic question of how we need to understand, to ‘this is helpful,’ to ‘let’s get this information into the hands of people who can use it,’” he said.

But like many other distributed projects, F@H has had its ups and downs. Its appearance on the PlayStation 3 was definitely an “up”—all you had to do was download the software and leave the PS3 on when you weren’t gaming. The console, powered by the way-ahead-of-its-time Cell processor, helped Folding@Home break the petaflop barrier in 2007, long before any supercomputer did. Millions of PS3s and PCs collectively achieved performance the top supercomputers could not equal.

Revival

At its peak, the PS3 brought in 15 million users, but it didn’t last. Sony ended F@H on PS3 in 2012. The numbers on other platforms fell more gradually, and by January 2020, F@H was down to 30,000 users. “Thirty-thousand is still pretty good,” said Dr. Greg Bowman, a professor at Washington University in St. Louis and a former student of Pande’s who now heads the project. “We had an interesting shift where people moved to laptops. The number of users was going down, but overall compute power was increasing because people participating had more powerful machines.”

One thing every distributed computing project has in common is that they are very processor-intensive, so participating PCs run very hot; your PC will go from 3 percent utilization to 100 percent very quickly, and if you are using stock cooling, the machine might overheat. So projects like F@H are best suited for a high-performance fan- or water-cooled tower PC, not laptop, where cooling is often minimal. That’s why the Mac is such a minor presence on the client list: MacBooks and iMacs are not suited to run at 100 percent CPU and GPU utilization.

Then in February, everything changed. Folding@Home suddenly went from 30,000 volunteers running the software in February to 400,000 in March—another 300,000 users came on board after that. There were so many users that the database ran out of potential simulations for them to crunch, and data coming in was so great that the servers were overloaded, said Bowman.

Despite these glitches, F@H zoomed to a peak performance of 1.5 exaFLOPs, making it more than seven times faster than the world’s fastest supercomputer, Summit, at the Oak Ridge National Laboratory. 

What caused this? For starters, interest in finding a therapy for COVID-19 helped. SETI@Home announcing the end of its project on March 31 also meant tens of thousands of people were looking for something new to run on their PCs. But the big boost came March 13, when Nvidia tweeted out a call to arms.

“Join us and our friends at @OfficialPCMR in supporting folding@home and donating unused GPU computing power to fight against COVID-19!” the tweet read. It linked to a Reddit forum with further details and information. “That helped a lot,” said Bowman.

A little too much, as it turned out. “We handled the first three- to five-fold increase in traffic without breaking a sweat, but when it hit twenty times, our servers were struggling to write all the info people were finding to disk,” he said. 

With just six servers at Washington University and partner sites at Sloan Kettering and Temple University, so much data was coming back and being written to disk that F@H stopped sending out work units. New servers have since helped them catch up. It did put a pause on the F@H team’s main project, a significant rewrite and update of the client app. That’s on hold for now, Bowman said.

Findings

What does protein folding have to do with COVID-19? The goal is to explore the folds of proteins on the coronavirus’s surface, trying to find spaces for other molecules to fit in a way that can interfere with the virus’s function.

The surface of the coronavirus could almost be viewed as a thing of beauty if it weren’t so deadly. It looks something like a Christmas tree ornament, with red appendages that look somewhat like an inverted traffic cone. These are called “spike,” a complex of three proteins. For the virus to attach and infect human cells, spike has to open. Bowman said the function reminded him and his group of the Demogorgon from the show Stranger Things.

Before spike opens, the interface it uses to interact with cells is buried internally; only after it opens can the interface be targeted by drugs. Bowman and his team programmed F@H to hunt for what he called “cryptic pockets” in this interface. “The idea is that if you look at the snapshots of what a protein looks like, it may or may not have pockets amenable to design. We’ve had previous success with antibiotic resistance and Ebola where a small molecule would bind tightly to shut off binding,” he said.

The project is still in the first stage, where researchers look for insight into how the spike proteins function and can be targeted. The project is doing computational screenings of molecules and will prioritize those that bind tightly to the virus’s proteins.

Bowman says he will share his findings with anyone who wants them. “We’re looking to put big data sets on the Internet for others to see how it works or pursue drug design,” he said.

Can it be sustained?

Folding@Home’s fall in users wasn’t an isolated incident. Once a novelty complete with teams (including a F@H team for Ars), it was one of many distributed computing projects that have fallen off in recent years. Kirk Pearson, who has maintained a master list of distributed computing projects for more than a decade, said the projects have lost their novelty. 

“When SETI@Home began in 1999, it attracted a lot of volunteers with its colorful screensaver and the new, exciting idea that anyone with a computer could find the first radio signal from an extra-terrestrial civilization,” he said. “But discoveries can take several years and a lot of repetitive work to find, and if the average person doesn’t discover something quickly, he loses interest and moves on to something else.”

He’s not exaggerating. GIMPS started in 1996, and in 24 years, it has found 17 massive prime numbers. The last prime number found by GIMPS was on December 7, 2018—it was 24,862,048 digits long. Before that, a 23.2 million digit prime was found in January 2018, and before that, a 22.3 million digit one was found in January 2016. That doesn’t exactly feed the need for instant gratification. 

Pande has another idea about what changed. “One thing that had a huge impact on distributed computing in general was the rise of cryptocurrency. It engaged a certain group of people and got their attention and had a similar impact on other programs like Folding@Home.” But, he adds, “I think people are realizing you can make a little money on cryptocurrency or have an impact on the world.”

Bowman hopes he can maintain this interest in having an impact. “I hope people will stick around once we get this virus under control and help us get control over diseases like cancer and Ebola,” he said. “We’re looking to learn why this is so much more infectious. For the foreseeable future, I think we will be quite busy with this. Over time, [we’ll] start putting it on other things.”

Pearson believes Folding@Home will inevitably lose a large amount of its new volunteers once COVID-19 is no longer in the headlines and the project ‘s novelty fades. But he’s optimistic many will stick around. 

“Every novel new distributed computing project creates a new group of excited volunteers, and many of those new volunteers will become long-term volunteers like me who will help Folding@Home’s other projects find cures for Alzheimer’s disease and cystic fibrosis; World Community Grid’s other projects find cures for childhood cancers, AIDS, and Tuberculosis; and many other projects make new biomedical and scientific discoveries,” he said.

As Pande put it, “We all play a part in infecting each other and can all play a part in hopefully curing each other.”