The U.S. Department of Energy’s Oak Ridge National Laboratory (ORNL) is now home to one of the first HPE Cray EX supercomputing deployments, which ORNL’s National Center for Computational Sciences is operating on behalf of the United States Air Force. Comprising two Cray EX supercomputers (formerly codenamed Shasta), the system — which became fully operational this week — delivers a combined 7.2 (peak) petaflops of numerical weather prediction capability to support Air Force and Army missions worldwide. The deal is valued at $25 million.

The Air Force named the new systems “Fawbush” and “Miller” after meteorologists Major Ernest Fawbush and Captain Robert Miller, who delivered the first tornado forecast at the Tinker Air Force Base in Oklahoma in 1948.

Each twin Cray EX supercomputer comprises 800 nodes, each equipped with two second-generation AMD 7742 Epyc Rome (64-core) processors. While currently boasting only CPUs, heterogeneity is on the horizon: each four-cabinet system can be expanded to 1,024 nodes — allowing the addition of future GPU-based nodes that would enable a 10X boost in floating point performance, according to ORNL. A small order of Nvidia Ampere A100 GPU-accelerated blades will be deployed this spring, and the additional cabinet space will also accommodate future-generation AMD+AMD nodes, i.e., the same node design as ORNL’s incoming exascale system, Frontier (using one AMD Epyc CPU and four AMD Instinct GPUs).

“We will absolutely have Frontier style blades in this machine,” said Jim Rogers, ORNL’s computing and facilities director, in an interview with HPCwire. Rogers, who led the acquisition, said the flexibility to have different node types really extends the research life of the machine. “If the research is extremely promising along any of these lines, that helps us determine where we want to go with a midlife upgrade,” he added, noting the project’s 10-year service life.

The deployment, internally called “HPC11,” uses 100 Gbps Slingshot networking with a standalone Slingshot fabric manager, which Rogers describes as being “unwrapped from the original Shasta design.” ORNL also separated all of the login and service nodes out into a separate hall. The original contract was announced in August 2019, one month before HPE closed on the Cray acquisition, so some of the subsequent system design decisions were made to accommodate scheduling risks associated with the acquisition process.

“We separated all of the non-compute, front-end stuff out very early, so that we could reduce the dependence on the Shasta software stack,” said Rogers. “And we basically rolled our own. Then as HPCM (HPE Performance Cluster Management) extensions became available to support the Cray EX, we integrated HPCM back into what we had initially started. So now it’s all completely supported configuration end to end.”

For their storage needs, ORNL went with a pair of high-performance Lustre file systems from DDN (SFA14K). “We did have to figure out how to tie those file systems back to the Slingshot fabric manager. That was actually one of the more difficult things, figuring out how to extend 100 Gig Slingshot fabric to the file systems. And we’ve managed to do that through some Arista networks switches,” said Rogers.

Given the mission critical nature of the system, resiliency and redundancy are serious design considerations. Fawbush and Miller each have their own unique power source with a dedicated power line. “We spent a significant amount of time on the design side working on the resiliency of this machine – the facilities, the power, the cooling and the system design – so that we can always be available,” Rogers said. “So, regardless of what’s going on, whether we have a scheduled or unscheduled outage, whatever the cause, there will always be more than enough computing capability available for them to get their work done.”

SchedMD Slurm workload manager runs in a federated configuration across the identical systems, such that they can be operated independently and concurrently. “Jobs are scheduled through Slurm and Slurm can run the job on either system; the user doesn’t have to know,” said Rogers. “The workload that the Air Force has can run completely independently on just one of the two halls, and they can still meet all their mission requirements. So for example, if Miller needs to go offline for whatever reason, scheduled maintenance, I can do that, and then Fawbush can carry the entire load. The federated Slurm capability kind of hides where the compute happens.”

As you might imagine, there are synergies with the Frontier project, which Rogers was happy to discuss. “The decision to move to HPCM with the standalone Slingshot fabric manager was driven by the ability to test it out at this scale,” he said. “Now Frontier is going to follow the same approach with the stand-alone fabric manager, HPCM and the entire ecosystem and stack. The experiences should dramatically improve the ability for us to start getting the Frontier cabinets in. Even though it’s a very different scale, the underlying experience is already there. And we have a lot of confidence in the underlying products: the software stack, the compute OS, the fabric manager, the HPCM implementation, all of that’s already well understood. So, the synergies are fantastic.”

Air Force weather researchers primarily rely on the Global Air Land Weather Exploitation Model (GALWEM), a commercial code based on the United Kingdom’s Meteorological Office’s Unified Model. In benchmarking using this production code, the new Cray EX system delivered more than six-and-a-half times the sustained performance of the researchers’ current system, Thor, a Broadwell-era HPE machine installed at the Offutt Air Force Base in Nebraska in 2016. The official transition to production on the new system is currently underway.

The additional computing power will immediately support finer model resolution — going from 17 kilometers to 10 kilometers between grid points. Such high-precision forecasting is critical for U.S. defense missions around the world.

“Looking to the future it is imperative that we take this step to ensure continued U.S. and allied dominance relative to our strategic competitors,”  said Col. Gary Kubat, acting Air Force Weather director. “This system has the growth potential to continue to meet emerging warfighter needs while enabling Air Force Weather to stay in lock-step with our modeling partners as we together develop the next generation weather models.”

The Cray EX deployment marks a major goalpost in a collaboration between the U.S. Air Force and Oak Ridge National Lab that commenced in 2018. The strategic agreement, made in recognition of the increasing scale and complexity of requirements for numerical weather forecasting, enables Air Force researchers to leverage the capabilities of ORNL’s National Center for Computational Sciences. ORNL provides their expertise in high performance computing facilities and infrastructure, systems administration and operations, and procurement and acquisition.

The win for HPE, which acquired Cray Inc. for $1.3 billion in Sept. 2019, continues the long succession of HPE and Cray systems installed at top weather forecasting sites worldwide, including NOAA and the UK Met Office. The new system is the first operational HPE Cray EX supercomputer at a federal facility, according to ORNL. The same architecture will power the United States’ first three exascale supercomputers, including the aforementioned Frontier, slated to be delivered at Oak Ridge by the end of the year.

“We are thrilled to have built the U.S. Air Force a new supercomputer that is one of the first operational systems powered by the latest HPE Cray EX supercomputer and managed by Oak Ridge National Laboratory (ORNL),” said Bill Mannel, vice president and general manager, HPC at HPE. “The end-to-end HPC technologies made possible by the HPE Cray EX supercomputer will enable greater speed and dedicated performance to advance simulations in weather forecasting that were never made possible before.”

ORNL provides the supercomputing services to the Air Force 557th Weather Wing, tasked with providing terrestrial and space weather information to the U.S. Air Force and Army.

In partnership with ORNL’s Computational Earth Sciences Division, the U.S. Air Force is on track to break new ground in weather forecasting. Targeted scenarios include:

Forecast[ing] stream flow, flooding, or inundation to predict how much of a given land will be submerged in water and the level of its depth. Researchers plan to achieve this by creating a global hydrology model that involves simulating hundreds of watershed and drainage basins to eventually increase accuracy in predicting future events.

Remote sensing of a cloud-covered area to address how to navigate impacted missions through forecasting the formation, growth and precipitation of atmospheric clouds. Researchers plan to achieve this by using comprehensive cloud physics that are not made possible with existing statistical regression models.

The new system was installed in two phases beginning in February 2020. Formal acceptance took place late January 2021, and a Certificate of Readiness was issued to the Air Force on February 5, 2021. As of February 8, 2021, Air Force Weather has initiated their transition to production.