We’re in Guangzhou, China, for the final round of the largest student cluster competition in the world. And to make it even more tense, there’s nearly $28k up for grabs.
The third annual Asia Student Supercomputer Challenge (ASC14) features sixteen university-backed teams from around the globe competing to see which team can wring the most HPC performance out of a 3kW power allowance.
The ASC14 competition is unique in several ways. The first is that the teams will be building their clusters out of the same building blocks. Vendor sponsor Inspur is supplying all of the gear, which makes it much easier (and less expensive) for foreign teams to make the trip. It also levels the playing field and puts a premium on the tuning and workload management skills of the students.
The road to ASC14 is a rigorous one, considerably more demanding than competing at SC or ISC. First teams had to register for the competition, which wasn’t all that rigorous. Then they had to complete the “preliminary round” of competition, which required them to submit their architectural proposal (hardware, software, interconnect, etc) plus a power consumption and performance evaluation.
The students also had to discuss the advantages and disadvantages of their proposed configuration. In addition, they had to optimize and run the four competition applications (discussed below) and provide the results to the committee. Inspur provided remote access to their systems so that student teams that didn’t have access to hardware could still compete in the preliminary round. Sheesh, that’s a lot of stuff to do.
The committee went through the submissions and invited the top sixteen teams to complete in the finals at Guangzhou.
Over a two-day period, students will be fighting to get maximum performance out of these applications:
- HPL (LINPACK) – a fixture in HPC and Student Cluster Competitions. It’s also the benchmark that determines placement on the TOP500 list.
- Quantum Espresso – a set of open-source codes used for nanoscale materials and electronic structure modeling. It uses pseudopotentials to approximate particular materials, which makes it easier for researchers to model combinations of materials, I guess. Speaking for myself, though, when I need pseudopotentials, I’m old-school – I’ll make ‘em myself.
- SU2 – a Stanford University developed open-source C++ code for Partial Differential Equation analysis and designing things that adhere to PDE constraints. Most commonly used in aerospace design, it’s an alternative to much more expensive proprietary packages. It’s the first time this application has been used in any cluster competition, and is probably new to almost all of the student teams.
- LICOM – the abbreviation of the LASG/IAP climate modeling system developed by China’s State Key Laboratory. It’s old-school Fortran 90, which might be a challenge for some students, requiring them to call their grandfathers for help. It is somewhat parallelized via MPI or OpenMP, but this hasn’t been tested thoroughly.
- 3D-EW (3D Elastic Wave) – code that is used to model how elastic waves propagates through an elastic, but uniform, medium. This application presents a special opportunity and challenge for the students. They won’t be running this one on their own clusters. They’ll be running it on 512 nodes of the biggest supercomputer in the world, Tiane-2, which is located just two floors below them in the Guangzhou Supercomputer Center. This will also require the students to utilize a whole passel of Intel Phi co-processors (assuming a passel = 1,536, which is three Phi’s per node.)
This competition combines the sprint aspect of ISC (where competitors only work on certain applications per day) and the marathon format of SC (where teams work on multiple applications simultaneously). Students will be working on LINPACK, Quantum Espresso, SU2, and the first round of 3D-EW (on Tiane-2) on Wednesday. Thursday will see the second round of competition on 3D-EW, LICOM, and a surprise application, which is, well, a surprise.
What’s at Stake?
Teams will be scored on how much performance they squeeze out of their hardware, relative to other teams, without topping the 3kW power cap. Teams can get 10 points by topping HPL, 15 points for winning any of the other applications, and up to 20 points for turning in the best result for the two rounds of 3D-EW on Tianhe-2.
Awards and cash prizes will be given in the following categories:
- Overall Gold Winner: 100,000 RMB ($16,000)
- Overall Silver Winner: 50,000 RMB ($8,000)
- Highest LINPACK: 10K RMB ($1,600 USD)
- Heterogeneous computing optimization: 10,000RMB ($1,600)
- Best applications performance: 10,000 RMB ($1,600)
There’s some serious cash on the line here, particularly for typically penniless college undergrads: about 180,000RMB in total ($28,900; £17,100), so we’re expecting to see the best effort possible from each team.
We’ll be posting much more content as the competition heats up, including team system configurations, video profiles of each team, and, of course, the results. Stay tuned!