NSF has awarded the National Center for Supercomputing Applications (NCSA) $10 million for its next supercomputer – named Delta – “which will kick-start NCSA’s next generation of supercomputers post-Blue Waters,” according to NCSA. Delta is a Category 1 award under NSF’s Innovative High Performance Computing Program for “Capacity Systems: production computational resources maximizing the capacity provided to support the broad range of computation and data analytics needs in S&E research.”
Asked if NCSA would run the LINPACK benchmark on Delta, Gropp said, “Probably – I don’t think it reflects computational science applications, but everyone expects it. We’ll run some other, more relevant benchmarks as well.” His comments reflect a growing consensus that LINPACK is less useful as an HPC benchmark as workloads and computer architectures have changed, and as AI (lower precision) computation techniques have become more widely used in science applications.
Few technical details were disclosed. Delta will have a heterogeneous architecture, feature more than 800 Nvidia GPUs, presumably the A100, and have “a nearly 10 petabyte high-performance storage subsystem, which incorporates three petabytes of flash-based memory.” The award requires acceptance by late 2021 so the timetable for standing up Delta is tight. NCSA director Bill Gropp told HPCwire he expects to finalize contracts with vendors by this fall and to start deploying Delta by mid next year.
“Since we headquarter XSEDE, we’re quite comfortable working with large groups of users,” he said. “The overall performance of this system for many applications is comparable to Blue Waters. But for things like machine learning workloads, Delta will in fact be faster than Blue Waters reflecting the evolution of the technology. That’s why we called it Delta. Certainly there have been lots of applications that have been taking advantage of GPUs.”
Here’s the official description of Delta taken from the award:
“The Delta system will be a unique addition to the NSF-funded ecosystem of advanced computing resources due to its balanced mixture of a substantial GPU capability alongside more traditional CPU processing capabilities. Taking advantage of next-generation processor architectures and NVIDIA graphics processors, Delta will include what will be, at its launch, the most performant GPU computing resource in the National Science Foundation portfolio with its 800+ graphics processors.
“The compute elements of Delta will employ an advanced network interconnect, delivering at least 100 gigabits per second of bandwidth to each compute node for application communications and access to an innovative storage environment. Data-intensive research will be accelerated by a nearly 10 petabyte high-performance storage subsystem, which incorporates three petabytes of flash-based, relaxed-POSIX file system which NCSA will use to demonstrate the effectiveness of modern file systems in advancing the performance of data-intensive and high-performance computing applications.”
NCSA was one of the early adopters of accelerators, noted Gropp, although Blue Waters was not a strictly speaking heterogeneous architecture. “We had the PS two cluster so, there certainly have been a lot of computational scientists who have taken the GPU plunge and taken advantage. But there’s also a lot of application groups that have held back for one reason or another, you know, one being that for some applications is it can be pretty easy to make the transition and for others not, you know, it’s harder. I think we’re well-positioned given our experience with GPUs with a mixed configuration machine that Blue Waters was (22,636 Cray XE6 nodes and 4,228 Cray XK7 nodes with Nvidia GPUs).
“One of my internal metrics for success is not that we simply provide GPU cycles to the applications that need it,” said Gropp, “but that we grow the community that can take advantage of machines like this.”
Gropp said there are no plans to retire Blue Waters which was stood up in 2012. “Blue Waters is pretty old but there’s capabilities that are rarely matched, even today among the biggest systems, like 1.4 petabytes of RAM, more than a terabyte per second of IO bandwidth and so forth. Blue Waters is old, but it’s still churning out great results.” He added there is plenty of space to bring in the new machine.
The move to a “relaxed-POSIX file system” is notable. Gropp said, “I think HPC has focused too much on POSIX compliance as an IO system. Part of the problem there is that for the application people, POSIX compliance means that they don’t have to change their code. Unfortunately, for somebody who’s actually implementing POSIX, it means a lot more work. POSIX has very, very strict consistency requirements.
“In fact, [the consistency requirements] are so strict that some HPC centers turn them off. They don’t actually provide a POSIX file system. They provide a big file system that provides the POSIX syntax, essentially the API, but isn’t fully correct. Turns out most applications are good with that. Some are not. And the Big Data people don’t use POSIX and have looked carefully at the consistency semantics that they need for their actions. One of the things that we’ve been looking at with our partners is a high-performance system [in which] you don’t need to change your application code interface, but in a way that is very precise, and relaxes the consistency requirements.”
Given that Delta will be a capacity system it’s expected to handle a wide variety of domains. “We do have a number of applications that we put in our proposal. We’ve also got things in new areas, and we’re doing a lot of work in digital agriculture right now. There’s also digital archaeology. There’s, of course, the current crisis that affects our choices. There’s some great opportunities for COVID-19 response and other planning.
“One of the ways we’re addressing having so many applications is partnering with the Science Gateways Community Institute (SGCI). So GPU-enabled gateways are going to be an important part of this. We also have an accessibility group in our College of Applied Health Sciences here at Illinois to help us look at accessibility issues for the interfaces,” he said.
No doubt there is some sting for NCSA in losing the big track one award to TACC last year (Frontera) which had a $60 million price tag. The Blue Waters contract was around $188 million. Except for exascale systems which are DoE-funded and not NSF systems, the new reality seems to be more modest sized machines. Time will tell. The $10 million size of the Delta award is actually the top amount allowed under the program and is similar to what Indiana University was awarded for its Jetstream 2 cloud system announced yesterday.
The Delta project is led by Gropp as principal investigator. He’s joined by co-PIs and NCSA colleagues, Gregory Bauer, senior technical program manager, Brett Bode, assistant director for the Blue Waters project, Timothy Boerner, deputy project director for XSEDE, and Amy Schuele, interim associate director of integrated cyberinfrastructure.
In the official announcement, Gropp is quoted saying: “NCSA is excited to continue its tradition of innovation in advanced computing with the Delta system, which will bring new capabilities to the nation’s advanced research computing community. With Delta, we will help emerging research areas, such as computational archaeology and digital agriculture, take advantage of new computing methods and hardware while simultaneously making advanced computing systems more usable and accessible to a broad community of researchers.”
NCSA will integrate Delta into the national cyberinfrastructure ecosystem through the Extreme Science and Engineering Discovery Environment (XSEDE), an NSF-funded virtual organization that coordinates the sharing of advanced digital services. Integration into XSEDE allows Delta to leverage the substantial portfolio of services and support offered therein and together deliver unprecedented advances in researcher productivity. This collaboration will promote synergy among multi-site workflows that include campus, national, and commercial cloud resources.