September 28, 2011

NetApp Gets First Petascale Supercomputer Win

Michael Feldman

NetApp flexed its newly acquired supercomputing muscles this week when it announced it would be supplying one of the largest Lustre storage system in the world for the Sequoia supercomputer to be installed at Lawrence Livermore National Laboratory (LLNL) next year. NetApp’s E-Series storage, which they inherited when the company purchased LSI’s Engenio business, will be used to provide 55 petabytes of disk arrays for the 20-petaflop Sequoia machine.

Multi-petabyte storage hooked up to elite supers is, more often than not, supplied by DataDirect Networks (DDN), the HPC vendor that has dominated the upper echelons of supercomputing storage for years. DataDirect currently is behind 60 percent of the top 50 most performant supercomputing systems in the world, supplying storage for 19.5 petaflops of HPC machinery. With this one Sequoia deal, NetApp has managed to reach parity, at least FLOPS-wise.

Sequoia’s precursor at Lawrence Livermore, the Blue Gene/P supercomputer named Dawn, uses DDN storage, as does the older Blue Gene/L system there. The L machine has 400 TB of disk and an aggregate I/O bandwidth of 40 GB/sec.

NetApp’s incursion into DDN’s territory could be a indication of a more competitive HPC storage landscape at the high end, or it could just be a one-off deal peculiar to this particular setup. As a storage component provider, Engenio is certainly no stranger to petabyte-level supercomputing. But NetApp acquisition’s of LSI’s Engenio storage division for $480 million this year placed a top tier storage company, with a large sales network, behind the company’s HPC hardware.

With $5 billion in annual revenue, NetApp certainly has the storage gravitas to stand behind such the super-sized Sequoia deployment. As NetApp CEO Tom Georgens said back in March when the Engenio acquistion was announced, “We can exploit this technology in a way the seller could not,” referring to LSI.

When it’s installed in 2012, Sequoia may well be the most powerful supercomputer in the world. The IBM Blue Gene/Q machine will consist of 96 racks, comprising more than 98 thousand compute nodes, and 1.6 million cores. Going by the latest Green500 ranking of the protype Blue Gene/Q, not only will Sequoia be a top performer FLOPS-wise, it will almost certainly be among the most energy-efficient supers on the planet.

The machine will be primarily used for running supercomputer simulations for the Stockpile Stewardship Program, the initiative that supports the safey and reliablity of nuclear weapons under the direction of the National Nuclear Security Administration (NNSA). Since live nuclear weapons testing is no longer permitted, the NNSA relies on these simulations to maintain the nuclear deterrent — applications that have become increasing performance-demanding as the weaponry has aged. Sequoia will also be employed for a range of other of scientific and engineering research that required elite supercomputing.

The orginal description for Sequoia called for just 1.6 PB of storage, which seemed a bit miserly for such a large amount of compute, so it’s only a little surprising they opted for another 50 petabytes. Neither the DOE lab or NetApp has commented on what the immense storage is to be used for, other than to state the obvious need for extreme I/O performance and scalability for the 20-petaflop behemoth.

The NetApp announcement did not specify which E-Series product is being used for the Sequoia deployment, but it might end up being based on the new E5400. In June, the company unveiled the E5400 storage line, making a point of its support for the Lustre file system and being benchmarked in the Lawrence Livermore Hyperion test bed. The E5400 is able to pack up to 60 drives and a petabyte of capacity in 24U, certainly making it suitably dense for a petascale setup.

The 55 PB of storage and 1 TB/sec of I/O bandwidth is a step change for supercomputers at LLNL, or nearly anywhere for that matter. Another Blue Gene/Q system, the 10-petaflop Mira supercomputer is slated for 70 PB of disk storage and 470 GB/sec of I/O. That system is scheduled to be installed at Argonne National Laboratory, also in 2012

But neither Sequoia or Mira will be at the top of the storage heap, capacity-wise at least. Last month, IBM announced plans to build a 120 PB storage system using the company’s home-grown General Parallel File System (GPFS). The customer and deployment date for that storage installation was not made public.