German KIT v Fighting Seawolves in student cluster deathmatch
Reg bookie reports from the digital ringside
HPC blog According to bettors, the upcoming ISC’12 Student Cluster Challenge in Hamburg will come down to a competition between home-country team KIT (Karlsruhe Institute of Technology) and the New York-based Stony Brook University Fighting Seawolves. (No, I don’t know what a ‘Seawolf’ is either.)
The competition next week, jointly sponsored by the ISC, HPC Advisory Council and Airbus, is the culmination of months of effort on the part of the five student teams representing universities from around the world. Their goal is to design and build a cluster that can out-perform all others on a varied set of HPC benchmarks and applications.
They’ve been working for months to figure out the right hardware/software combinations and learn how to make the applications run like greased weasels. The only limit to their designs is a hard power cap of 13 amps – everything else is wide open. More details on the overall competition are here.
This ISC competition is a European version of the Student Cluster Competition, which has been a staple at US-based SC shows for the past five years. I’ve become a big fan. It’s both gratifying and fun to watch the students work so hard, learn so much, and then thoroughly enjoy the fruits of their labour during the week of the competitions.
To infect others with my enthusiasm, I’ve been handicapping the teams and running (virtual) betting pools to see where reader sentiment lies. KIT (team profile here), the home team, saw a lot of early action on both the LINPACK and "overall win" categories. The German team is rated as a 3-1 favourite among bettors on both.
But are KIT bettors basing their wagers on cold, hard facts or on less objective factors? It’s difficult to say for sure – except in the case of one bettor who noted “my son is with them”. Very touching, brought a tear to my eye, in fact. But objectively, the KIT team has its work cut out for it. It is a first-time competitor and the students are going up against much more experienced clubs. As the home team and the only European team, at what point will the students start to feel the crushing pressure that comes with being the sole repository of an entire continent’s hopes and dreams?
Stony Brook bettors make it pretty clear in their comments that they’re betting with their hearts. (American team profiles are here.) And they bet their hearts out in the final hours of our wagering pool, moving their beloved Seawolves from 5-1 to a 3-1 favorite.
Their comments included “Go Stony Brook!” and “Stony Brook University is home of the best and the brightest,” plus the simple and eloquent “STONY BROOK!” But can this team live up to the high expectations of its fans? Sure, they’ve been to the big dance before, participating in two SC Student Cluster Competitions. They won the big prize in 2009, but finished out of the money in 2010’s New Orleans bout.
China’s lead team, from Tsinghua University in Beijing, slipped a bit in late betting to finish at 7-1 odds of winning it all and 8-1 for posting the highest LINPACK score. This is probably a serious underestimation of Tsinghua’s chances in Hamburg. This team has been tested. They had to fight their way past 30 other universities in a six-team intra-China tournament to become one of two teams to carry the red flag. After four days of clustering and benchmarking, Tsinghua U was the overall winner. That’s significant. (You can see profiles of both Chinese teams, and details on the intra-China tournament here.)
To win the top slot in Hamburg, Team Tsinghua topped NUDT (National University of Defense Technology). NUDT competed at the SC’11 Student Cluster Competition in Seattle and just barely came in second to a very experienced team from Taiwan. NUDT brought a highly sophisticated, GPU-heavy system that ran rings around other competitors on the HPC applications. They also competed in the intra-China play-in round, finishing second to Tsinghua. NUDT as an 8-1 and 11-1 underdog is a bet that I’d take all day and is probably, along with Tsinghua, the best value on the board.
The Colorado team is clearly seen as an underdog by our bettors, coming in as a 16-1 and 11-1 long-shot to win overall honours and LINPACK respectively. But I’d argue that they’re being overlooked. This is a team that has more experience in these competitions than the rest of the field combined. They’ve been to every big-time cluster competition; they have a solid and experienced lineup, and a dedicated coaching staff that’s committed to winning.
The Buffaloes have seen some success over the years, taking the LINPACK award in ’09 and winning the Fan Favorite award in 2010. They’ve put in the effort and done the work necessary to take the big prize – it just hasn’t happened yet. Could this be their year? Might we see a victorious Colorado team hoisting the first ever Kluster Kup above their collective heads?
We’ll know soon. The LINPACK part of the competition kicks off on Monday, with final scores turned in at the end of the day. Students will get the data sets for their application runs on Tuesday and Wednesday, with final results due at 4pm sharp on Wednesday. The winning teams will be announced at the awards ceremony just an hour later.
For those of you who can’t make the trip to ISC in Hamburg to watch the competition in person, I’ll be covering the action live and will provide frequent updates. Check back often to see how your favourite team is faring. ®
A power limit of 13 Amps? Just up the input voltage get as much power as you want.
Re: Stoney Brook?
Same here. No joy with said video (...It's Vid>AY<oh...)
Blocked cos I'm in the wrong country... I didn't need a passport to get into Basildon but it seems I'll need one to get out.
Such is life.
Yes and it also depends on whether those are "average" 13 amps, "directed average" 13 amps or RMS 13 amps, or whether a sinoidal form of the current is required.
It seems like you could get quite a bit more power out of a socket if you only were to consume power in non-standard and bad for the power network ways.