ISC cluster smackdown: Students whip out their tools
White Intel bread with GPU sprinkles
HPC blog The final system configurations for the ISC 2012 Student Cluster Challenge have been locked down and can now be revealed. For the ISC inaugural event, we’re seeing a lot of sameness but a few key differences.
On the "sameness" side, every team is sporting various flavours of Intel Sandy Bridge processors, with speeds ranging from KIT’s low-powered 1.8GHz E5-2650’s to Stony Brook's 2.9GHz E5-2690s. Everyone is also running some flavor of Linux, with most opting for Red Hat.
You’ll notice that I appended the term ‘max’ in several of the columns. This is because the chart below represents the gear the students brought with them – not what they’ll necessarily be running or the speeds they’ll be using.
The only limitation in the contest is power: there’s a hard cap of 220 volts at 13 amps. And it’s a truly "hard" cap. At the SC version of the event, the student teams could vary above their power cap for a minute or two and woulkd receive a warning to crank things down. At ISC, they simply can’t get any more power – just like the real world, as competition organisers point out.
While all of the teams have tested their systems at home, they needed to fine-tune their configs to take into account the power they have at the event location. So even if they thought they had brought a huge number of cores that can run at high frequencies, it’s almost certain that they’ll need to either disable nodes or run them at slower speeds in order to stay under the power limit.
It’s interesting to peruse the table and note that there are significant differences from team to team. KIT and Stony Brook have more cores than anyone else, but went for slower CPUs and doubled up on the memory. At the last SC version of this competition, almost half the teams used GPUs as accelerators. It seemed to pay off; the GPU-enabled configurations took four of the top five finishes.
But here at ISC, only China’s NUDT team is using GPUs in their system. In talking with the teams, they all (except for NUDT) said that the apps didn’t seem to be a good fit for GPUs. This is pretty much what the non-GPU teams said at the SC competition back in November – again, except for NUDT and, of course, the eventual winner, Taiwan’s Tsinghua University.
It’ll be interesting to see how this all plays out. Will high core counts make the difference? Will NUDT’s GPU-centric strategy pay off? Will the competition come down to hardware power or how well the students handle the software configuration and optimisation? Watch this space. ®
Sponsored: Hyper-scale data management