There can be only one SC13 student cluster champion team. And it is...
Building big iron against the clock for fame and, well, fame
HPC blog Team Longhorn from the University of Texas won the Overall Championship Award in the Big Iron division of the SC13 student cluster-constructing competition in Denver.
Now that the dust has settled on the contest – which challenges undergraduates to build and benchmark HPC rigs against the clock – here's the lowdown on the results. (If you're new to these competitions, you can find out more here.)
Big Iron: Texas Repeat
Team Longhorn ... Students from Texas with their, er, gongs
The only thing more difficult than winning the Overall Championship in a student cluster competition is doing it twice in a row. Texas has become only the second team to score a repeat win in a major international tourney, joining Taiwan’s National Tsing Hua University, which took the gold home in 2010 and 2011.
This is Team Longhorn’s fourth cluster competition since 2010. Along the way, it notched a couple of historic firsts. In 2010, it became the first student team to break through the TeraFLOP barrier on the way to winning the LINPACK crown.
At SC11 in Seattle, Team Longhorn was the first team to use liquid cooling for its rig. The group immersed its entire system in a vat of mineral oil. It was very messy and earned the crew style points, but it didn’t give the team enough added oomph to win an award.
Texas came to Salt Lake City in 2012 with a new attitude and a handful of GPUs. The team used both of these to grab its first Overall Championship award and a celebratory dinner at the Cheesecake Factory.
Texas scored strongly in both application performance and the interview portions of the competition. The Longhorns were the top finisher in four of the seven scoring categories and took second place in an additional two. Its biggest single win was on the HPCC benchmark suite, where it topped other competitors by about 25 per cent.
Hardware-wise, the Texans brought in what they call a "balanced system". By this, they mean that they didn’t go hog-wild on accelerators, stocking four NVIDIA K20s (one per compute node) as opposed to other teams that fielded up to eight GPU crunchers.
Team Longhorn didn’t have the most CPU cores either. Its five-node cluster (four compute nodes plus a head node) hosted 100 CPU cores with "only" 320 GB of RAM – but which totaled 3.2GB per core, which is also less than other teams.
Who won what and how?
Mass effect ... Team Boston at SC13
Team Boston (the Mass Green Consortium, aka Team Chowdah) took second place in the race for the Overall Championship, finishing about 15 per cent behind the Longhorns and 10 per cent above the third place finisher.
Boston scored a big win in the round to run and benchmark a "mystery app", which turned out to be OpenFOAM, topping all other teams by a significant margin. They also took second place on HPCC, GraphLab, and NEMO5.
Team Chowdah had six nodes containing twelve Xeon processors giving them 120 CPU cores, with 768GB of RAM (6.4GB per CPU core). They also had six NVIDIA Kepler K20 GPUs – one for each node.
The gang's only weakness, if it could be called that, was its middle-of-the-pack finish on the interview portion of the competition.
The team from the Friedrich-Alexander-Universität, of Erlangen and Nürnberg in Germany
Team Germany, aka Team Friedrich-Alexander-Universität (FAU), took home the third place honors, which is pretty good for a first-time competitor.
On the hardware side, the crew brought four nodes, each with dual Xeon processors. They had a total of 64 CPU cores and 512GB of RAM, yielding 8GB per CPU core. They also had a grand total of eight NVIDIA K20 GPUs to spur things along on the GPU-friendly apps.
The students scored outright wins in the interview category and a big win on GraphLab. The team also took third place on HPCC and OpenFOAM. A nice debut for the Germans.
Some of the other Standard Track noteworthy results include Team Venus 2.0 (University of the Pacific) taking second place on two of the WRF categories and a third place finish on LINPACK – which shows great improvement over its results last year.
The Aussies from iVEC finished in the money on WRF and were in the thick of things on the other applications as well. Team Buffalo from the University of Colorado also put up a good effort, scoring in the middle of the pack on all competition categories. However, they were busy tracking rocket launches to Mars (team members were responsible for tracking telemetry from the probes), so can be excused if their full attention wasn’t on the competition.
China’s NUDT was also in the race until the very end. As everyone already knows, NUDT took home the LINPACK crown with a score of just over 8 TFLOPS. But they also turned in a good performance on the other competition applications, taking the third spot on NEMO5, one of the WRF categories, and in the interview section.
Team Knoxville had a string of bad luck at the worst possible time. The basic problem was that their competition system has some serious (and difficult to pinpoint) hardware incompatibilities. The only solution to their problem was open-heart surgery, meaning an interconnect transplant. This cost them a whole day to solve out of a two-day event.
While the surgery took up precious competition time, there was another cost as well: they lost all of their application optimization modifications in the process. They spent weeks crafting these mods and their system performance wasn’t nearly the same without them. It’s sort of like trying to compete in an auto race with a stock minivan.
But they persevered and finished the competition, and that’s what counts here. They didn’t give up and throw in the towel. It’s too bad that their hardware betrayed them. Judging by their interview scores, the team had pretty good application and HPC knowledge – enough to have made a very good showing in the competition – if only their stupid hardware had cooperated.
However, that’s how it goes in the highly competitive world of big-time Student Cluster Competitions. It’s not all science and numbers – luck (bad or good) plays a role.
Congratulations to all of the teams who competed at SC13 this year. We’ll probably be seeing NUDT at the upcoming Asian Student Supercomputer Challenge next spring, and will definitely be seeing some of these schools competing at the newly enlarged ISC’14 Student Cluster Challenge this summer in Germany. ®
Sponsored: Hyper-scale data management