On February 4, the new 500-plus teraflop “Ranger” supercomputer at UT’s Texas Advanced Computing Center (TACC) officially went into production and is now running a number of applications for U.S. researchers that received some of the early allocations. The system was funded by the National Science Foundation (NSF), which awarded $59 million to TACC ($30 million for the system and $29 million for operation) for the first Track 2 petascale system. Track 2 is aimed at high-end, but sub-petaflop systems for the U.S. science research community. At 504 petaflops (peak), the Ranger system is now the most powerful system on the TeraGrid and one of the top supercomputers in the world.
Developed by Sun Microsystems, Ranger was built from 3,936 Constellation blade servers, each containing four quad-core “Barcelona” Opterons running at 2.0 GHz. The 15,744 Barcelona processors in Ranger come from the batch that suffers from the highly publicized translation lookaside buffer (TLB) problem. TACC has used the recommended Linux kernel memory patch to work around that particular problem. Reportedly, the patch has little, if any, impact on performance.
The blade servers are connected using two Sun DDR InfiniBand “Magnum” switches, which were designed specifically for the highly scaled out cluster architecture of the Constellation line. The switch contains 3,456 ports, allowing the Sun system designers to collapse the hierarchy of switches and reduce the cabling normally required to connect thousands of nodes by a factor of six.
Deployment of Ranger has been something less than smooth. The NSF Track 2 funding for the TACC supercomputer was awarded back in September 2006, with the original machine slated for production in the middle of 2007. That system was to use the older socket F dual-core Opterons and top out at 105 teraflops. It would then have been upgraded to 421 teraflops with the quad-core Opterons toward the end of 2007. (When the award was made, the Barcelona chips were expected to be only fashionably late, leaving plenty of time for an upgrade.)
As it turned out, the more persistent Barcelona slippage didn’t matter. The Sun Constellation blades themselves were late, which killed the plan for the initial dual-core Opteron version of the system. There were also problems with some of the InfiniBand cabling, which caused an additional delay. According to TACC Director Jay Boisseau, to make up the lost time without a system, they worked out a deal with Sun to increase the size of the system for the four years of operation. So instead of having a 105 teraflop system for six months and 421 teraflops for three and half years, Sun would deliver a 504 teraflop system that would be in operation for all four years.
The upside for TACC is that the new system required an additional Magnum switch because additional nodes were needed to reach 504 petaflops. This gives TACC quite a bit of headroom to build out the system yet further with additional nodes. Theoretically, one could build a petaflop system if somewhat faster quad-core Opterons show up (say 2.5 GHz) just by hooking up additional blades into the existing cluster. That, however would require a significant new funding source. In the short term, it’s more likely that TACC will consider using some of the unused InfiniBand ports for storage nodes and add another petabyte or two of disk capacity to the existing 1.7 petabytes. The TACC folks have also shown interest in upgrading their visualization resources. Hooking a new visualization cluster onto the Ranger would make a lot of sense for a system that is expected to be doing a lot of big science.
In its current configuration, Ranger is poised to take on some of the most computationally-intensive science research in the U.S. Expected applications include QCD (quantum chromodynamics), earthquake modeling, climate modeling/weather forecasting, material science research, astrophysics, CFD and turbulence codes, bioinformatics, chemistry, physics and geosciences, to name a few. Since the system and its operation are funded by the NSF, any researcher engaged in open science at a U.S. institution can submit a request for Ranger cycles. Although a U.S institution has to be involved, the researcher need not be a U.S. citizen. Even scientists outside the States can gain access to Ranger by collaborating with U.S.-based researchers.
Boisseau says one of the applications that scales particularly well on Ranger is POP (Parallel Ocean Program), a classic HPC code that models ocean circulation. According to him, POP ran better on Ranger than on their Lonestar system (based on 2.6 GHz Xeons), even without taking advantage of the additional cores available on Ranger.
“Lonestar is one of the most powerful compute platforms in the country and POP ran 35 percent faster on Ranger than on Lonestar, which really shocked us because Lonestar’s cores are really fast,” said Boisseau. “But [POP] just happened to optimize really well for the Barcelona processor.”
Not all of the early applications they’ve tried are running as well on Ranger. One thing they’ll have to do is optimize the system software for a 16-core blade configuration. Up until last year, most software was optimized for 2- or 4-core blades. With the more recent Intel quad-core Xeons, some people have started using 2-socket, 8-core blades. Most of the MPI libraries haven’t been optimized for that level of scalability yet. In particular, MPI collective operations like scatter, gather, broadcast and reduce should be optimized for the size of the node for greatest effect. TACC is currently working with a team at Ohio State to tune an MPI implementation for these larger node systems.
The compilers for Barcelona are also new, and not a lot of real-world application testing has been performed. TACC is making both the Intel and PGI compilers available to users. Other system software used by Ranger, like the Linux OS (CentOS distribution) and related drivers, may also have to be tweaked to improve code performance. Boisseau expects the first several months will see steady application performance increases as the underlying system software is optimized for the new Ranger platform.
Although application performance is what users are most interested in, Boisseau says they have run the Linpack benchmark on Ranger and are quite pleased with the result (which at this point they are keeping to themselves). With 504 teraflops of peak performance, once could expect the Linpack figure to be at least 350 teraflops, which would easily earn it the number 2 on the current TOP500 list. But by the time the next list comes out in June 2008, Ranger may find a number of newer supercomputers jockeying for the top spots.
Ranger is expected to be officially dedicated in a ceremony at TACC on February 22. Sun Microsystems CEO Jonathan Schwartz is expected to be part of the festivities.