The Discover system at NASA’s Center for Climate Simulation was designed with scalability and flexibility in mind, starting with its original nodes in 2006 and consistently evolving and growing with new hardware to handle the computational and data demands of complex long-term climate models.
In total, with the individual scalable units, the supercomputer is a combination of mostly IBM racks with some Dell and SuperMicro in the mix, culminate in just over a petaflop at peak across the over 43,000 cores.
But for this center, it’s out with the old and in with the new for around one-quarter to one-third of the system each year. The spare nodes are shipped off to be used in other research projects, leaving Dan Duffy, High Performance Computing Lead for NASA and his team with a revolving chain of decisions. For this round of system upgrades, they moved away from their IBM-heavy approach and looked to SGI’s Rackable line for memory, power, and CPU balance.
In the latest upgrade round, the team removed 1200 nodes from Discover or about 14,000 cores—the equivalent of roughly 150 teraflops of compute that was installed in 2010. Now, in the same footprint and power envelope, the new 18 racks they are receiving from SGI offers more performance (over a petaflop peak) and 30,000 cores for 1000 nodes. This upgrade represents almost an 8x performance boost in the same space when compared to what was available four years ago.
The key to these efficiencies are inside SGI’s Rackable cluster line, which for NCCS means 1.9 petaflops of capability powered by the Xeon E5-2696 v3 processors fed with a non-blocking fat tree FDR interconnect from Mellanox. Duffy said that while it was tempting to hold off and wait for EDR, especially since these applications are between 30-40% message passing, they needed the system up and running right away.
Further, Duffy added that if they were going to delay to have EDR in the machine, there are some key technologies coming in the future that are also worth the wait. While he and his team have been experimenting with both GPUs and Xeon Phis for some time, there’s still a great deal of code work to be done. He said that they’re keenly interested in where Knight’s Landing will take their workloads and are taking a particular shine to 64-bit ARM, if for no other reason than what it will mean for the competitive processor ecosystem.
“A couple of years ago we explored the Phi and made good progress but our results show that its performance is the same we can get in a node for this application. With GPUs, we upgraded to the NVIDIA K40, but the numbers are low. Our goal is to keep developing the applications to exploit manycore and hybrid core architectures. We’re interested in getting our hands on Knights Landing and comparing that to Haswell and eventually Broadwell. In a separate project we do a climate model on Power systems as well.”
When one takes into account the workload across this scalable architecture at NASA, it’s clear how important the CPU to data to storage balance is. The main application for complex long-range climate modeling is called GEOS-5, which integrates multiple models using the Earth System Modeling Framework—itself a beast in terms of computational requirements and complexity. Duffy says that to date, teams have been able to scaled past 30,000 cores. While right now they’re just running this at NCCS on roughly 8,000 cores, the plan is to take to its full heights with the SGI system. “It’s no small task,” laughed Duffy. “We are simulating the climate for a long amount of time at very high resolution. On top of that, there’s a 2 year run of this to look at the next generation of earth observing instruments to validate the work happening with that. That 2-year simulation is at a 7 km resolution and has generated 5PB of data.
This means that aside from raw compute and the critical FDR interconnect backbone, memory and storage are other considerations. In addition to the benchmarks that NASA uses for these applications to test everything from message passing to generalized performance, the memory footprint needed strong consideration. Duffy said that while he certainly isn’t at liberty to endorse any particular vendor, the SGI benchmark results delivered across all of these areas for this particular application area, opening new doors for the company, which just rolled out yet another line of HPC systems targeted at other workloads and featuring next generation Xeon E5-2600 v3 processors as well as a new UV system for data-intensive workloads that is based on the Xeon E7-8800 v2 processors, the capability to snap in GPUs and Xeon Phis, and most interestingly, the seventh generation of their NUMAlink technology which is arguably the true differentiating factor for SGI at the high end. We’ll be talking shortly with Dr. Goh about these new systems and their path forward with customers like NASA, who according to Duffy, are looking beyond standard rack systems and into a future that could include emerging technologies they’re watching on the processor, memory (specifically the HP Memristor, he says), and storage sides.
Computational story aside, at the end of the day, what Duffy and his systems are providing, are a glimpse into detailed, long-range climate narratives that will offer a first glimpse into how the planet’s climate has evolved from Paleolithic times until the present.