Tell me if you’ve heard this one before. IBM is planning to deliver one of the fastest supercomputers in the world, to help unravel the mysteries of the universe. Deja vu?
This latest IBM colossus will be a 360 teraflop machine that is headed for the University of Toronto and is to be shared by the SciNet Consortium, a group that includes the university and a number of research hospitals. Applications destined to run on the new super include the usual HPC suspects: aerospace, astrophysics, bioinformatics, chemical physics, and climate change prediction. The new machine will also be used by the CERN-run ATLAS project, which is investigating the forces that govern the universe.
From the IBM announcement, here are the main bragging points:
“[T]he machine is expected to be among the top 20 fastest supercomputers in the world; 30 times faster than the peak performance of Canada’s current largest research system. It also represents the second largest system ever built on a university campus, and the largest supercomputer outside the United States.”
The new system is slated to be fully operational next summer, although a partial implementation could be running as early as January. The acquisition is part of a five-year deal that is expected to cost $47 million.
One of the unconventional aspects of the system’s design is that it will incorporate both Power6- and Intel Nehalem-based clusters using IBM’s new iDataPlex platform. According to company sources, 300 teraflops will be Nehalem-based nodes and 60 teraflops will be Power6-based. This is yet another example of IBM’s increasing comfort with hybrid computing platforms. IBM’s more famous Roadrunner system combined AMD Opterons with Cell processors to deliver the world’s first petaflops computer. The idea behind hybrid computers is to give users the ability to run code on the most appropriate hardware in order to speed execution time.
Of course, the difficult part of hybrid computing is the software. Initially, individual applications will execute on one architecture or the other. Cluster Resources’ Moab cluster scheduler will be used to map jobs and will also include specific enhancements for a stateless, diskless multi-architecture cluster. The entire cluster is hooked to a storage system and uses GPFS as the backend file system. Both processor architectures will share these resources.
The plan is that over time, as software is optimized for the heterogeneous environment, it will be possible for applications to execute certain stages on the Nehalem portion and other stages on the Power6 portion. IBM says the entire cluster is modular and will support the changing of the OS and architecture dynamically.
The announcement of the Canadian super comes on the heels of IBM’s deal with the UK Met Office to deliver a 125 teraflop supercomputer for weather forecasting. That procurement represented a five-year contract that would have the Met Office system topping out at 1 petaflop by 2011. Unlike the Canadian machine, the UK system is pure Power6, as are the two 145 teraflops systems ordered by the European Centre for Medium-Range Weather Forecasts (ECMWF), the 125 teraflop system acquired by the Max Planck Society (MPG), a 76 teraflop system installed at the National Center for Atmospheric Research (NCAR), and the 60 teraflop machine running at SARA, in Amsterdam. That’s over 600 teraflops of Power6 deployed — or scheduled to be deployed — since the beginning of 2008. If you ignore the Roadrunner machine as an anomaly, Power6-based supercomputers acquisitions are outpacing the much more storied Cell-based supers.
In a way, that’s a little surprising. The sporty little 4.7 GHz Power6 chips tend to run hot and are among the least efficient processors in the increasingly important performance/watt metric. It’s no coincidence that the Power6-based clusters are all water cooled. So why is this architecture so successful? Maybe it’s because the IBM sales team could sell central heating in Singapore. Or maybe it’s because there’s a lot of Power-based software out there expecting great single-threaded performance and ready to hop on to the next processor generation. Whatever the reason, it will be interesting to see what the Power7 chips bring to the table. Hmm… I’ll bet they’ll be used to build one of the fastest computers in the world, to help unravel the mysteries of the universe.