National University of Defense Technology, China
Team Tianhe looks for Seattle glory
SCC team profile China’s National University of Defense Technology (NUDT) is the same organization that brought to life the 2.56 PetaFLOP Tianhe-1A supercomputer. This system, with more than 14,000 Intel Westmere processors (186,368 cores) and 7,168 NVIDIA Tesla GPUs, shocked the world when it came out of nowhere to take the top slot on the Top500 list in November 2010.
NUDT is an interesting organization. They are a driving force – maybe the driving force – behind China’s burgeoning HPC infrastructure. They’re the first to develop Chinese supercomputers at the 1, 10, and 100 GFlop levels plus, of course, China’s first PetaFlop system.
It’s also the only HPC organization supported by the China Ministry of Education and the China National Science Foundation Innovation Group. And now they’re sending a team to the SC11 Student Cluster Challenge.
So what are they bringing to the competition? On the hardware side, they’re sporting servers from China’s own Powerleader, meaning some flavor of Intel processor. I think it’s safe to assume that they’ll be going with a hybrid CPU-GPU system along the lines of Tianhe. On the software side, they’ve selected Red Hat as their o/s.
NUDT won’t be the only team packing GPUs this year. Other teams have used them in the past and still more will be using them this year. It’s beyond debate that GPUs, used properly, can take number-crunching to whole new levels of performance.
In some cases, hybrid systems with CPUs and GPUs can crank throughput by 5 – 50x. They also provide way more flops/watt than general purpose CPUs, which is all good for SCC teams who get only 26 amps to play with.
Like prescription medicine, GPUs aren’t a panacea. Popping GPUs into a system willy-nilly is like downing a handful of brightly-colored capsules some dude gives you at a rave party; it doesn’t necessarily result in a fun time. Last year, the GPU-wielding teams didn’t realize much benefit from them. They did have good performance on some apps, but it wasn’t enough to push them to the top of the standings overall.
I think that the GPU-packing teams failed to make the grade last year mostly because of how they were using them. I don’t think that the apps they were running were properly optimized for GPUs, and none of the teams had a lot of experience in GPU care and feeding.
This is where Team China might have a leg up on the competition. They have the team that built Tianhe behind them and, I think, plenty of chances to use Tianhe for testing their code. While the other GPU-fueled teams will have more experience this year than last, they don’t have the same access to a large hybrid cluster – or the experts who use it on a daily basis.
I don’t have a lot of information about Team NUDT from a personal standpoint. Their team looks to be composed mostly of computer science undergraduates, with a couple who are concentrating on networking. They’re looking forward to the experience and are confident that they’ll do well – a common trait among this year’s competitors.
What’s different is that they’re backing up their confidence with some numbers. Last year, we had three teams (barely) hit TeraFlop performance. Without giving too much away, I can say that Team NUDT expects to improve on this by a considerable margin.
This Chinese team is a bit of a mystery. It’s their first SCC, and no one knows how well-prepared they are or how they’ll hold up under the pressure. However, they have a lot of built-in advantages and should definitely be considered a top contender.