SCC team profile In five years of Student Cluster Challenge (SCC) action at the annual Supercomputing Conference, Purdue has yet to hoist the championship trophy over their collective heads. (There isn’t a championship trophy, but there should be.) They’ve had an impact, winning the Green Award for most flops/watt in 2008 and 2009, but they haven’t taken either the LINPACK or the overall crown. (There should be a crown too.)
This year, Purdue returns a team anchored by veterans who have participated in, or at least observed, the competition. They will be one of the more experienced teams in Seattle this year; the majority of team members say that they know their way around HPC. This could be bluster, but if their application is any guide, these guys know their clusters.
They’re a confident bunch – maybe because they’ll be sporting their trademark sledgehammer, or maybe due to their experience. Either way, almost every team member said that they are one of the top teams or that they expect to win. You have to like the attitude, particularly after five years of coming up a little bit short.
Purdue will be riding Intel hardware and Red Hat Linux in the competition, with Portland Group compilers. One significant difference from last year is that the team expects to get their competition hardware much earlier this time around, giving them more time to get familiar with and optimize the system.
Purdue is another team that’s chock full of computer science majors, which could hinder them when it comes to dealing with the specialized applications they need to run in the competition. Recognizing this potential weakness, Purdue reached out to students and professors in non-computing scientific domains.
Their plan was to find graduate students in the application fields and convince them to enroll in the special ‘Team Purdue SCC’ course. These students could then help the team learn about the science behind the code and, hopefully, how to optimize the target apps.
It’s hard to tell whether this extra effort has paid off. When asked how they felt about the various workloads in the challenge, the team was highly confident that they’d turn in a good LINPACK – which is what you’d expect from an experienced team of computer science majors.
But when it comes to the subject matter apps (PFA, GAGET, MrBayes and POP), they’re not as sure of themselves. However, this could be due to typical Midwestern U.S. modesty, which would preclude them from coming right out and saying, “We own this one” or “We’re confident.” That would be bragging, and Midwesterners usually leave that stuff to the Texans. When asked how they’re going to cope with the lack of sleep during the two-day marathon SCC session, they cited division of labor (which allows more team members to sleep), coffee, and power naps as a key coping mechanism. One student said:
In the past, I have competed in robotics tournaments on about 90 minutes of sleep and as a CS major sleep is always a premium. As long as there is excitement and I have something to keep me busy lack of sleep won't be an issue. (During the wee hours of SCC, when the systems are just clicking away, teams learn to make their own excitement. One team last year converted their booth into a theater and watched movies during down time.)
The Purdue team has already learned some valuable lessons, according to the members…
Preparing for this challenge has helped me realize that good documentation can make all the difference. One of the apps was particularly lacking in this area, and made me try to document what I do better. (If this experience convinces just one person to document their code better, then it’s all worthwhile, right?)
It has been pretty interesting to look at the different scientific applications that scientists actually run on large computing resources and the different ways you have to approach getting them to run on different computing resources.
- To prepare our team for cluster challenge the university created a class to allow us to get more work with our applications in. Since the class is in the Earth and Atmospheric Sciences department we are also working with grad students who use some of the applications we are working with and utilize Purdue's clusters on a daily basis. This also gives us the ability to interface with people in other disciplines about our applications, what they use on our machines, and cluster challenge as a whole.
So maybe the Boilermakers have soaked up enough domain knowledge to give them a leg up on the other teams?
Last year, I thought Purdue was the most serious and “heads-down” of the teams I observed. When I’d stop their booth for updates, it was pretty clear that they were all business. I kind of felt like I was annoying them at times with my semi-coherent random chitchat. I can’t blame them; I wouldn’t talk to me much either, but I don’t have a choice.
Time will tell if the 2011 Purdue team has what it takes to compete and win in Seattle. Is this the year they break through to the top? Will they “Boiler Up” in 2011?
SC11 Student Cluster Comp profiles
- National Tsing Hua University, Taiwan
- National University of Defense Technology, China
- Tecnológico de Costa Rica
- Boston University
- University of Colorado
- Purdue University
- Nizhny Novgorod State University
- University of Texas