Weather forecasting has come a long way since June of 1977, when the European Centre for Medium Range Weather Forecasts (ECMWF) first contracted Cray to deliver one of its early Cray-1A systems across the pond. This was the first time a Cray found its way to the old country—an installation that set the stage for a number of new deployments of both vector and shared memory systems to power European weather prediction over the next several decades.
The first Cray system at ECMWF enabled the weather center to offer a 10-day forecast powered by a weather model that achieved sustained performance of 50 megaflops (against the system’s theoretical peak of 160 megaflops). These systems were followed by the Cray X-MP/22, then an X-MP/48, followed by the Y-MP 8/8-64, C90 (with a gigaflop of theoretical peak), and then into shared memory territory with the T3D. This was the last system ECMWF bought for a stretch in favor of Fujitsu and then Power-based systems from IBM. Now, 36 years after choosing their first Cray system, EMCWF is taking the supercomputing back.
The Big Blue machines that are being swapped out for the XC30 early this year were ranked at 51 and 52 on the most recent Top500. If you’re wondering why there are two systems of equal proportions that are essentially tied, it’s because specific operational requirements demand a two-machine approach for centers who provide model outputs that power the weather forecasting efforts of an entire continent—as is the case with ECMWF.
Isabella Weger, who heads the Computer Division at the weather center (and has been instrumental in the two-cluster approach decision that set the trend for other weather modeling centers worldwide) explained that having separate clusters in the datacenter offers more resilience for operational forecasts.” In essence, one system runs the center’s operational forecasts, which are the critical products they deliver to the 20 member states and 14 co-operative states in Europe that our models for regional and local weather forecasting.” The other cluster runs the center’s research workloads, which includes activities centered on improving their numerical weather prediction model and offering a more comprehensive view into atmospheric behavior.
While both clusters are busy chewing on their own workloads, all operational data is available to both machines. The dual storage clusters, which will now be Cray Sonexion-based systems, are cross-mounted across the compute clusters so EMCWF has access to the data readily available in the event that they need to restart the forecast during a system upgrades or problems.
Although Weger and team set the dual-cluster trend at ECMWF, this is a rather unique approach to continuity in Cray CEO, Pete Ungaro’s experience. As he told us, “we haven’t seen this kind of configuration outside of operational weather forecasting centers, really. Most people that are using our machines for research tend to build the single biggest engine they can. However, the operational requirements we see even in demanding commercial markets are not as evenly focused from an operational standpoint as what EMCWF and other major weather centers need.”
This dual-approach to cluster and storage scenarios is the direct result of Isabella and team’s need to ensure constant delivery of the critical forecasting models centers in Europe rely on. And the data’s importance doesn’t end there—EMCWF has an extensive tape library of model outputs from decades gone by which totals over 50 petabytes of historical climate data. Further, she says their system generates around 50 terabytes per day. These data are used by climate and atmospheric scientists around the world who require detailed data from outdated model output for advanced climate change and other longer-range atmospheric studies.
For now, however, it’s about adding more fine-tuned resolution to the models to better help governments prepare for weather events. “If you imagine a grid around the globe, our current model resolution is 16 km between grid points and our plan is in 2015 timeframe to go to a finer resolution of 10 km, hence the driver for compute resources.”
All of this takes some serious compute horsepower, which beginning early this year, will mean the use of the Aries interconnected Cray XC30 “Cascade” supercomputer with a multi-petabyte Sonexion storage system—again, split into two separate clusters. Ungaro described the environment as accelerator-free (although the system is capable and Weger said they are considering the future of accelerators for their application) noting that “each of these [Ivy Bridge] systems are in two different halls, each about 19 cabinets, about 3,600 nodes, all interconnected with our Aries interconnect, so about 80,000 cores in each of the machines.”
To put all of this compute into some context, keep in mind that over 60 million observations are factored into the overall forecasting model at EMCWF. It starts with observations, which come from a range of sources, many from satellites, others including ground based observational tools, buoys, and airplanes. These observations provide the baseline for the forecast.
“We take these many observations and process them to drive a base point for the atmosphere,” Weger explains. “These are all observations from different points in time and space, and we must snap these into a grid of sorts that spans the globe in the proper space and time.” This is EMCWF’s process of “data assimilation” which in itself is both data and computationally-intensive—and it all happens before the forecast model has begun.
Complex forecasting is not a “one-shot” system. Since no forecast is perfect, a sense of probability for weather events must also be calculated. “We run an ensemble of 51 forecasts per day, each with some changes in the initial conditions to get a sense of probability. If you relate this to a hurricane, for instance, the model gives you the projected track of the storm with different conditions.”
“It’s about performance, of course, but also very important are resilience and reliability and also, portability,” added Weger. She notes that they strive to keep their forecasting system portable across architectures so that with each procurement cycle they have many vendor choices. “The application is mainly Fortran and whenever we optimize or develop code we try to make sure it doesn’t inhibit us from making architecture choices–we don’t want to be locked into a specific vendor or architecture.”
While Weger didn’t comment on their experiences using the IBM Power architecture, she and Ungaro both agreed that the benchmarking and procurement process was lengthy and detailed. EMCWF has a scientific and operational 10 year strategy that defines the upgrades they do across their model (called the Intergrated Forecasting System, which is the code comprises the model and data assimilation). Much of their upgrades are driven by the need for a lot of computing resources to power increases in model resolution, thus allowing the center to use more observational data and offer a better representation of the physics in the atmosphere in the model itself.
Adding more computational power to the forecasts makes quite a difference over time. While it might not sound like much in passing, the ability to add one more day of quality forecasting per decade, could make an incredible difference during potentially severe weather events. “A seven-day forecast today is as accurate as a 5 day forecast was 20 years,” explained Weger.
Ungaro, who was in the room during our chat with Weger, was beaming by the end of the conversation when the topic went back to the “full circle” nature of this new system at ECMWF. “We are very proud to have this kind of history and to help provide the systems that can save lives and make such a difference in the world,” he said.
While we might be able to do some speculative math on the potential placement of the new Cray system on the next Top 500 list—and its ability to provide more power for the models than the IBM Power-based system, time will tell. We’ll check in on this story again once the system appears on the June list.