The United States’ National Oceanic and Atmospheric Administration (NOAA) last week announced plans for a major refresh of its operational weather forecasting supercomputers, part of a 10-year, $505.2 million program, which will secure two HPE Cray systems for NOAA’s National Weather Service to be fielded later this year and put into production in early 2022. The long runway gives the managed service provider, CSRA (a General Dynamics Information Technology company), about a year to get the equipment in place, configured and accepted, and then from February of 2021 to February of 2022, NOAA will transition its code base over from the current systems.
With this hardware upgrade, ongoing model enhancements and NOAA’s emerging Earth Prediction Innovation Center (EPIC), NOAA says the United States is keeping pace with other leading weather forecasting centers around the world. The prominence of the U.S. weather forecasting capabilities has at times been called into question, perhaps most notably when U.S. models stumbled while forecasting Hurricanes Sandy and Harvey.
The new supercomputing deployment represents a tripling of operational computational capacity for the U.S. weather forecasting agency.
Each identical Cray Shasta system spans 2,560 dual-socket nodes — housed in 10 cabinets — powered by second-gen AMD Epyc ‘Rome’ 64-core 7742 processors, connected by Cray’s Slingshot network. The total system memory per machine is 1.3 petabytes. Cray’s ClusterStor systems provide 26 petabytes of storage per site (a flash storage system with 614 terabytes of usable space and two HDD file systems with 12.5 petabytes of usable storage).
The peak theoretical performance of each Cray system is 12 petaflops, which combined with NOAA’s research and development machines brings the agency’s aggregate operational and research capacity to 40 peak petaflops. Shasta systems haven’t hit the Top500 list yet, but at a ballpark 80 percent Linpack efficiency, they’d be looking at a 25th place ranking on the current (Nov. 2019) list. As always though — and no more so than for weather prediction and storm forecasting — the only thing that matters is real-world performance. HPCwire spoke with some of the NOAA/NWS HPC team about what international leadership means to them.
“You can imagine there are a lot of different ways you can measure leadership,” said Brian Gross, director of Environmental Modeling Center for NOAA’s National Weather Service. “[You can] measure it by hurricane track, accuracy of the upper level flow, surface temperature anomalies… it really depends on what your application is. [Regarding] how good the model is, we’re always compared to some of the [leading] centers worldwide. And we actually work pretty closely with the other worldwide operational centers. We have scientific exchanges with the European Center, for example. So the idea that we’re in a fierce competition is kind of a weird one for us as we work with these folks on a pretty regular basis.”
Housed at GDIT-managed facilities in Manassas, Virginia, and Phoenix, Arizona, the new Crays will replace eight smaller machines that comprise a heterogeneous mix of processor and cluster types. Moving to a unified architecture will streamline NOAA’s operations, while maintaining the weather center’s primary-plus-backup workflow (more on that below).
The outgoing equipment includes older IBM iDataplex gear, a pair of Cray XC40s (Luna and Surge), that were deployed in 2016, and a pair of Dell systems (Mars and Venus) installed in 2018. The agency is currently adding additional Dell machines to update the iDataplex systems so they are maintainable for the final two years of the managed service contract (with IBM).
Recall that NOAA’s operational centers are still managed by IBM, which procured the Cray and Dell systems after its x86 business was transferred to Lenovo in 2014. That IBM contract is up in February of 2022, at which time, GDIT will take over.
The transition to a new managed service provider coincides with a change in filesystem technology. After about 20 years of being on GPFS, NOAA is switching its systems over to Lustre. The move should not be seen as reflecting NOAA’s preference for a given filesystem, rather the agency provided the specification for performance-based requirements for the contract and what it required in terms of availability (99 percent system availability) as part of the open bid process and let industry decide what the best fit was in terms of the total proposed solution. “We were essentially looking for what the best fit was for what the integrator could provide…[and] the best performance-per-dollar with the availability requirements that we require for operational use of the system,” David Michaud, director, Office of Central Processing for NOAA’s National Weather Service, told HPCwire.
The decision to go with homogeneous x86 systems was made in a similar manner. NOAA asked the integrator to provide the best solution on the benchmark codes it utilizes. Meanwhile NOAA is exploring GPU technology on its research and development systems, and keeping its options open for the next hardware procurement. The contract with GDIT (there’s an 8-year base with a 2-year optional renewal) is split into two periods. The first task order covers the two Cray CPU-based systems, but the second period is still undefined, affording NOAA time to explore and assess the realm of possibilities as technology develops and as leadership computing facilities, many of which have moved or are moving to heterogeneous GPU-powered systems, help develop and influence technological advancements.
The twin Cray systems are perfectly symmetrical between geographically-segregated sites (Manassas, Virginia, and Phoenix, Arizona), and take turns acting as the primary or backup system. Michaud explained that on any given day, NOAA can run at production, its full operational 24×7 modeling suite on one of the systems. The backup system is used for transition to operations and other development work while it’s not in use as the primary, and NOAA can switch the orientation of the primary and the backup site in operations within a 15 minute period, and does so regularly, at least on a monthly basis.
The arrangement assures redundancy, as data is always mirrored to the backup system, offering advantages from a troubleshooting and maintenance perspective and providing an added layer of protection for the mission- and safety-critical work of weather prediction. “If we make a change to one, we know we can test it, and then we can apply the change to the back-up system as well,” said Michaud. “We know if one’s not behaving similar to the other, we can identify the differences and troubleshoot them. And then, the other thing that’s really important is for the type of work that we do, given storm systems and other weather systems can be massive in scale and encompass hundreds of miles, it’s really beneficial for us to have the separation of the sites, so that if we have any issues on one site, we can switch to the other site.”
The significant supercomputing upgrade targets three separate areas for model improvements: resolution, complexity and the size of ensembles. “We want to go to higher resolutions that would capture the finer-scale features in the phenomena we’re predicting,” Gross told us. “We want to create and implement more comprehensive models to include as much of the complexity that is going on in the atmosphere as we can in our models — can we improve our model physics, for example. And then the last piece is growing the size of ensembles that we use, which give us a lot of information in how certain we can be in any particular forecast; ensembles inform our level of confidence in the numerical guidance that we produce. All of these areas are going to be improved when we move on to the system.”
Last summer, NOAA upgraded its deterministic global forecast system, and its next big upgrade will be the ensemble system. Currently, NOAA is incorporating the new dynamical core that it put into the deterministic Global Forecast System (GFS) into the Global Ensemble Forecast System (aka the GES). “We’re aligning the ensemble system now with the deterministic system — and part of that is complexity and ensemble size. We’re looking forward to increasing the ensemble size of the GES so that we can get better information on forecast services,” said Gross.
NOAA’s contract with GDIT has a total estimated value of $505.2 million, spanning a base period of eight years with a two-year optional renewal. GDIT provides its supercomputing resources as-a-service through NOAA’s Weather and Climate Operational Supercomputing System (WCOSS) contract. The value of the first task order, written under the larger contract, is $150 million dollars to provide managed services over five years.