The tenth annual Student Cluster Competition kicked off Monday at the SC16 conference here in Salt Lake City, Utah. This is a competition for the ages, with some wildly divergent hardware configurations, more teams than ever before (14), plus some special features that are sure to throw the budding HPCers for a loop.
What is a Student Cluster Competition anyway? Glad you asked. It’s an event where teams of six undergraduate students, from universities around the world, build their own supercomputers. They then race their computers to see which teams’ system can run a set of HPC benchmarks and scientific applications the fastest. The only rules are that the hardware has to be available on the market and the teams can’t go over the power limit of 3,000 watts.
The students have to find vendor sponsors (who supply the hardware), configure their systems, and build out whatever software infrastructure they need to run the applications. They also need to learn as much as they can about the apps in order to understand how to optimize them for their particular hardware.
Grueling marathon is grueling
Like the subhead says, the SC competition is tough. Students will be clustering for 48 hours straight, striving to wring the most performance on the applications without going over the power cap. They’re on their own during this – no outside coaching allowed.
There are four main applications on the menu this year.
HPL (LINPACK): Students will be running the same benchmark that is used to rank systems on the semi-annual TOP500 list. It’s a measure of how quickly a system executes floating-point operations to solve a dense system of linear equations. Every cluster competition in the world uses LINPACK as their first benchmark – it’s tried and true.
HPCG: This is a newish benchmark from the same guy who brought us LINPACK (Dr Jack Dongarra). This one is a torture test that is designed to more closely match today’s applications in HPC. You can find out more here.
Distributed Password Auditing/Recovery: Students are given data sets that contain one-way hash passwords. Their task is to verify the passwords by hashing out the hash and coming up with the actual password. They can use their own techniques, but most will use some sort of comparison to a dictionary of words or word permutations to come up with the right answers.
Paraview: This is a visualisation program that allows researchers to quickly build visual representations of their data. It’s used in a wide variety of scientific fields and is equally adept at showing both quantitative and qualitative data. It also supports 3D, which is all the rage these days. Student competitors generally agree that Paraview is the most challenging of the applications this year.
ParConnect: This is the first time that students in a cluster competition will be working to replicate results from a scientific paper from a past SC conference. In this case, it’s assembling the genes of the huge number of organisms sampled from pond scum. They won’t be working with actual pond scum, but if I had my way…
Mystery Application: GROMACS is a fully featured molecular dynamics simulation suite designed to study proteins, lipids, polymers, and the like. If you have to analyse molecules, GROMACS is for you. But it also has wide applicability in other chemical and biological research.
There is another twist to this year’s event. In the first year of the competition, there was a power failure in Reno (site of SC07). Student clusterers were thrown into a tizzy as their machines suddenly went kaput. The teams that checkpointed their work were able to recover relatively quickly, but the teams that didn’t lost hours of effort. The organizers of the competition this year are going to flip a switch at an undisclosed time to replicate that experience. That should be a lot of fun to watch.
Next up, we’re going to profile the fourteen teams and handicap the field while taking a look at their individual configurations. There are a LOT of surprises in the configurations this year, so be sure to take a look.