PSC’s Big Data and AI Supercomputer Replaced by New Bridges-2 Platform
From the vastness of neutron-star collisions to the raw power of incoming tsunamis to the tiny, life-and-death details of how COVID-19 progresses, the Bridges platform at the Pittsburgh Supercomputing Center (PSC) has seen it all. Now Bridges has taken its final bow, ceding the title of PSC’s flagship high-performance computing (HPC) system to the larger, more advanced Bridges-2.
Funded in late 2014 by the National Science Foundation (NSF) as a component in the NSF’s XSEDE cyberinfrastructure “ecosystem,” the $10-million Bridges—also supported by the NSF with an initial $10 million in operational funding—has, among numerous other advances:
- powered artificial intelligence (AI) that beat the world’s best human poker players
- identified the anonymous printers who put on paper fundamental 17th Century works on individual liberty that underlie the U.S. Constitution
- improved predictions of severe weather to lengthen warning times
- shed light on the genetic resilience of species to climate change
- offered gene researchers an easy-to-use tool to assemble the largest DNA and RNA sequences
Bridges was so successful that, in 2018, the NSF extended its life with an additional $1.9 million for an extra year of operations.
“Bridges was our first and highly successful step in creating a new computational platform for accelerating new and novel forms of computational research,” said Shawn Brown, Bridges’ principal investigator and director of PSC. “Bridges-2 will take the next step in accelerating rapidly evolving data and AI research with greater capabilities and the same excellent team supporting the research.”
Recently Bridges answered a call to service, enrolling in the international COVID-19 HPC Consortium. This made the system available to scientists doing COVID research via a unique overnight approval process. As part of the Consortium, Bridges (along with the D.E. Shaw Research Anton2 supercomputer at PSC) supplied over $880,000 worth of computing time and storage to scientists searching for public health, drug and vaccine remedies to prevent and treat disease caused by the SARS-CoV-2 virus.
Like all of PSC’s systems, Bridges was available for open research or educational purposes at no cost to users, as well as at cost-recovery for private companies. In total, Bridges powered 2,100 projects by 16,000 users at 800 institutions of science and learning across the U.S. and the globe.
The Predominance of Memory—and of Helping the Newbies
Bridges’ first became a twinkle in the eyes of PSC’s technical staff and leadership when the center’s NSF-funded Blacklight (2010-2015) found intense demand meeting scientists’ needs in two key HPC areas.
The first was memory. Memory—the same as RAM in a personal computer—was not traditionally a priority in HPC design. Raw processing speed (in flops, or floating-point operations per second) stemming from many small-memory but fast processors had been the need in earlier HPC problems such as aircraft design or weather prediction. But the advent of new applications for HPC systems, including but not limited to assembly of massive DNA sequences such as in the Human Genome Project, created a need for a computer that could store “Big Data” ready for its processors without necessarily having the fastest or largest array of processors. Blacklight offered RAM in two massive chunks that proved popular with scientists in these “new research communities.”
These new communities posed another challenge for PSC’s staff. The scientists in these fields had never before used HPC and didn’t need or want to learn how to program supercomputers. So Blacklight also featured an ease of entry that few HPC systems of its generation could match.
In designing Bridges in cooperation with Hewlett Packard Enterprise, PSC doubled down on Blacklight’s strengths and added more. Bridges would be an even bigger Big Data resource, with regular, large, and extreme memory nodes tailored for different types and sizes of projects, and an even larger data storage system, called Pylon. The 276-TB-RAM Bridges (276,000 gigabytes, compared with a “hot” personal computer of today’s 64 GB) was no slouch in speed either, offering 1.3 petaflops (roughly 12,000 times as fast as the hot PC).
“Big machines are like old friends, it’s always sad to see when they go,” said Olexandr Isayev, assistant professor of chemistry at the Mellon College of Science. “We used PSC Bridges very productively for a number of years … It was a fun experience, and we are looking forward to Bridges-2 for even more fun—but most importantly, to solve even more challenging problems in chemical sciences!”
Bridges also offered even more user-friendly operating modes, ranging from direct programming for the HPC veterans to “virtual machine” and “interactive” modes for the newbies. The system proved so easy to use that it helped power PSC’s STEM education efforts, and was used by students from the middle school to graduate levels. Bridges’ use included 130 such educational allocations.
Bridges went one step further by containing graphics processing unit nodes (GPUs, the same as graphics cards in personal computers). GPU technology greatly speeded many types of computation, and powered the explosion of AI successes beginning in 2012. The 21,056-core “heterogeneous computing” Bridges gave researchers the ability to route portions of their computations to whatever type of processor would speed their work the most. PSC staff wrote most of the software that made that vast and speedy data-traffic routing possible.
In 2019, the NSF funded an expansion of Bridges’ AI capabilities with the latest GPU technology. Bridges-AI, which features the world’s first NVIDIA DGX-2 system for open research, fueled even more and more sophisticated AI work on Bridges. Bridges-AI will continue this mission, as a part of the new Bridges-2.
While Bridges’ final disposition is still being worked out, its components will not go to waste. Some of them will be repurposed into Bridges-2. Others will be used for research and philanthropic purposes by other organizations.
The $10 million (plus $10 million in initial operational funding) Bridges-2, funded in 2019, will in turn build on Bridges’ successes, providing even more massive AI and Big Data capacity to serve both scientists for whom HPC has long been a standard tool and those in new communities. About three times larger than Bridges, with 64,512 cores, Bridges-2 has completed its early user period and begun regular operations.
Later this year, Bridges-2 will be federated with PSC’s new Neocortex, an AI-specialized HPC system also funded by the NSF last year.