SGI’s entry into the multi-petaflops sweepstakes began last week with its announcement of the Space Act Agreement. The agreement outlines a multi-year partnership between SGI, Intel and NASA, with the goal of deploying petascale supercomputing capabilities for the NASA Advanced Supercomputing (NAS) facility at the Ames Research Center. The three organizations will work together on a project called Pleiades to develop two systems: the first, a 2009 system with a capacity of one peak petaflop; the second, a ten petaflop system for 2012.

Pleiades refers to the constellation of the seven sisters, a cluster in the constellation of Taurus. The Pleiades metaphor applies not only to the likely cluster architecture of the new supercomputers, but also to the seven organizations involved in the project: the three development partners (SGI, Intel and NAS) and the four main mission directorates at NASA (Aeronautics, Exploration Systems, Science, and Space Operations) who will end up using the new systems.

NAS is tasked with a number of high profile computing projects, including space shuttle simulations, climate modeling and future space vehicle design. The future petascale capabilities are seen as the computational foundation of NASA’s ongoing missions of space exploration and terrestrial research.

According to Dave Parry, senior VP and Product GM at SGI, the Space Act Agreement is a statement of intention, not a commitment to any specific system procurements. Any resulting contracts will be dependent upon NASA funding and mission priorities over the next four years, which, given the structural deficits in the federal budget and a new administration in 2009, are likely to evolve substantially over time.

The Pleiades collaboration between SGI and NASA is the result of a long-term relationship between the two organizations built over many years. The space agency was the first customer to deploy large SGI Origin SMP systems. Later, when SGI developed the shared-memory Itanium-based Altix machines, NASA became one of the early users of that technology. The current 88.9 teraflop Columbia supercomputer at the Ames center is a supercluster of two generations of Itanium-based Altix servers. The NAS division also owns a Xeon-based 43.5 teraflop Altix cluster, dubbed RTJones. On occasion, NAS has dabbled in IBM gear (a 4.8 teraflop Power5-based cluster, called Schirra), and systems from the now defunct Linux Networx, but the agency has remained unusually loyal to SGI for its major deployments.
 
“I see the Space Act Agreement not as a point-in-time event, but rather as part of the long-standing partnership that we have,” Parry says.

But the two systems being envisioned under the agreement have very different development paths. The 2009 petaflop machine will likely be an evolution of SGI’s NAS Technology Refresh supercomputer, also announced last week, and slated to be installed this summer. That system, a 20,480-core Altix ICE supercomputer, is designed to provide more than 240 teraflops of computing power. Parry says that the current ICE technology based on Intel Xeon processors “certainly provides expandability to systems up to and beyond a petaflop.” With Intel’s Nehalem processor technology and 40Gbps InfiniBand just around the corner, a one petaflop ICE cluster in 2009 should be within reach.

The ten petaflop system will be something else entirely. The conventional wisdom in the industry is that getting to multi-petaflop systems (and ultimately to exaflop systems) is going to require manycore processors, system architectures that are capable of scaling those processors, and integrated system software adept enough to effectively utilize all those resources. Parry notes that “a simple scaling of the socket and node count of the 240 teraflop system that we’re installing would not lead you to a 10 petaflop system … that would be impractical from a physical perspective. Clearly, we will need to do something new.”

Neither SGI nor any other HPC system vendor has commercial examples of such architectures. Cray and IBM are probably the furthest along the multi-petaflop path, thanks in part to support by DARPA, via the HPCS (High Productivity Computing Systems) program. SGI didn’t make it past the first cut of HPCS, so the company’s partnership with Intel and NASA should serve to bolster some of that lost R&D.

The partnership with Intel is particularly interesting, inasmuch as Cray just recently teamed up with the x86 chipmaker to collaborate on research and development of multi-petaflop systems for the same post-2010 timeframe. It is certainly conceivable that the SGI and Cray efforts could end up converging on the same manycore Intel processor architecture.

The fact that NASA is sticking with SGI gear points to the agency’s desire to follow the path of least resistance to petaflops. The challenge for SGI and Intel is to deliver commercially-viable petascale supercomputing that can support NASA’s HPC application software. That probably means a machine powered by x86, Itanium, or some combination of the two, and a software platform that makes the shift to manycore relatively painless.

“This is not going to be a stunt with a bunch of cell phone processors to get a high peak number,” says Parry, “It is really about deploying relevant production-ready capability.”