Researchers are growing an open supply, mixed-precision benchmark instrument for supercomputers

As Frontier, the sector’s first exascale supercomputer, was once assembled on the Oak Ridge Management Computing Facility in 2021, figuring out its efficiency in mixed-precision computations remained difficult. This hole in figuring out was once no longer an oversight, however reasonably an indication of the way new supercomputer programs that excel at combined precision will stay in computational science, which has been ruled by means of double-precision-focused programs for the majority of its historical past.

Double-precision – or 64-bit – mathematics is the main usual for mathematics accuracy in simulations. Blended-precision computations—16 or 32 bits—regularly computed by means of GPUs can give you the ranges of accuracy required at a lot upper speeds, particularly for knowledge science and synthetic intelligence. However there’s no broadly to be had open supply instrument for trying out lower-resolution efficiency on a big scale on GPU-accelerated supercomputers, which first noticed standard deployment in 2012 with OLCF’s Titan machine.

Thus, researchers at DOE’s Oak Ridge Nationwide Laboratory have advanced a brand new cross-platform benchmarking instrument package deal simply in time for Frontier’s Might 2022 release: OpenMxP. They’ve additionally made it to be had to different computing amenities as open supply code.

“The supercomputer is a crucial basis for keeping up technological management in computing in the USA, and we’re pushing the limits of supercomputing — pun meant — however you’ll’t give a boost to it if you’ll’t measure it, which highlights the significance of benchmarking,” stated Feiyi Wang, “This reference implementation of OpenMxP as an ability usual will get advantages all different vanguard computing programs,” stated Team Chief for Massive-Scale Synthetic Intelligence and Analytics Strategies, or AAIMS, on the Nationwide Heart for Computational Science at ORNL.

Because of its application in supercomputer analysis and operation, OpenMxP was once just lately identified as a finalist for the 2023 R&D 100 Award within the Tool/Products and services class.

Run the numbers

OpenMxP implements the HPL-MxP benchmarking job, which was once presented in 2019 and is the business usual for measuring mixed-precision benchmarking for supercomputing programs. The HPL-MxP items an issue to be solved—a dense machine of linear equations—however no longer the instrument to resolve it. That is as much as requirements. Prior to now, for the OLCF Summit supercomputer, its GPU chip distributors advanced and ran particular codes to guage their velocity in appearing mixed-precision calculations.

“Previously, distributors or integrators have all the time controlled this type of usual. They increase their very own code — it is their secret sauce to tell apart themselves and supply their distinctive aggressive benefit. They arrange it, and we take their effects ‘as is,'” Wang stated.

This was once no longer an possibility for Frontier as a result of it’s powered by means of AMD’s CPUs and GPUs, so new benchmark codes would should be advanced to run correctly at the next-generation chips. When NCCS Director Georgia Taurasi raised the query of whether or not ORNL may just run the usual itself, Wang recommended forming a crew to take action. In April 2021, they started learning the benchmark drawback itself and consulting students who’ve labored on identical issues.

“We had no enjoy fixing this sort of drawback, so we bumped into issues of instrument applications that we did not await. You simply be expecting a message-passing interface to paintings as you assume it’s going to paintings on your head. “We had no enjoy with it,” stated Mike Matheson, technical director of the OpenMxP venture on the AAIMS staff. But it surely did not paintings that manner.” “So, we might check out issues, after which it would not paintings, after which we might communicate to people, after which we might check out one thing else. We have been roughly exploring the long run, attempting to determine what in reality labored. “It was once a finding out curve, and we simply needed to do it.”

Thankfully, the Frontier venture was once nonetheless a number of months clear of of completion. Sadly, this additionally approach they’re going to be growing code for a tool they have not been in a position to check it on but. However as soon as OpenMxP was once able for preliminary rollout in mid-2021, the crew had some other very speedy — if no longer as speedy — machine close by to function a check mattress.

“Our purpose was once a Frontier machine, however Frontier at that time did not in reality exist. So, we took good thing about what we had, which was once Summit,” Wang stated. “It was once in fact lovely excellent, which means that after we had a solid machine first of all, were given our code up and operating, and set Summit at that scale, our code was once battle-tested. We knew it could scale, and The remaining was once to evolve or get ready it for Frontier.

In Might 2022, Frontier was once able to release. So was once OpenMxP. Frontier’s preliminary mixed-precision benchmark of 6.86 exaflops — or 6.86 billion floating level operations according to 2nd — places it on the best of the 2022 HPL-MxP listing. 12 months later, it reached 9.95 exaflops to take the highest spot as soon as once more. The Ecu Prime Efficiency Computing Joint Undertaking’s LUMI supercomputer extensively utilized OpenMxP to post its HPL-MxP utility and got here in 2nd position to Frontier within the June 2023 ratings.

Multipurpose instrument

Hanging a bunch within the festival of entries isn’t the true energy of OpenMxP. The instrument package deal in the end supplies insights into how neatly GPU/CPU supercomputers carry out, serving to to give a boost to their efficiency by means of revealing how small adjustments in programming can result in leaps in computational velocity. With the short, robust effects it unearths, OpenMxP too can show to computational scientists the benefits of the use of GPU-equipped programs able to appearing mixed-precision calculations.

“A large number of simulations are fixing huge programs of equations, and generally all of those are simply double precision, the place researchers take a hammer and pass down the trail they know,” Matheson stated. “It made sense up to now as a result of there have been no special-purpose units.” . . “However with the appearance of all GPUs doing low-precision calculations sooner than CPUs, it has enabled this new answer procedure to be horny.”

Additionally, OpenMxP can function a device to resolve some issues in science and engineering with unparalleled speeds and effort potency. The usual HPL-MxP drawback solved by means of OpenMxP is composed of enormous linear programs of equations, which might be additionally the construction blocks for science and engineering packages.

In 2022, the ORNL crew — together with NCCS analysis scientists Wang, Matheson, Hao Lu, and Jens Glaser — used OpenMxP as an answer for TwoFold, a instrument suite that predicts how strongly a given drug molecule will bind to pathogens and which predicts the 3-d impact. The construction of the way it is connected to the objective. TwoFold has been named a finalist for the 2022 Gordon Bell Particular Prize for HPC-Primarily based COVID-19 Analysis by means of the Affiliation for Computing Equipment.

“The true factor we are seeking to do is transfer science ahead by means of giving scientists a work of open supply instrument that they may be able to construct on and regulate to resolve their clinical issues. And because we are at this kind of huge scale and scale, OpenMxP can assist cope with the most important clinical issues in some way that sooner”.