Tsinghua University topped all other competitors at the ISC18 Student Cluster Competition with an overall score of 88.43 out of 100. This gives Tsinghua their second gold medal of the year, their first was at ASC18 this spring. It also gives them a chance to win the coveted Triple Crown of Clustering if they win at SC18 later on this year.
The top three finishers represented the cream of the student clustering crop, with Nanyang Technological University in second and perennial front runner CHPC in third. Honorable mentions go to newcomers ShanghaiTech and Heidelberg for finishing fourth and fifth respectively.
ShanghaiTech also wins the coveted Fan Favorite Award, given out to the team that gets the most votes in our online poll. We had nearly 4,000 registered votes and ShanghaiTech pulled in the largest total, which gets them the trophy. Good job ShanghaiTech.
How Did They Win?
Let’s dive into the individual scoring and see what’s what.
First, as we covered in earlier articles, Germany’s own FAU hammered down the Highest LINPACK Award while The New Pride of Shanghai, ShanghaiTech, took top HPCG honors. South Africa’s CHPC burned through the top HPCC score like using liquid oxygen to start your backyard BBQ.
Unfortunately, the benchmark section of the competition is only worth 10% of the overall score collectively. So while winning the benchmarks is a great start, benchmark wins alone won’t get a team to the Overall Championship. If you want the big points, you have score big on the applications.
An unheralded first time team from Brazil, UFPR, shocked the power teams by taking first place on the Grid application. Grid is a C++ mathematical object library that’s designed to exploit parallelism. Students were presented with a dataset and required to process it in the least amount of time possible. The UFPR win is even more impressive when you factor in the fact that they aren’t driving a fancy cluster festooned with GPUs. The Boys from Brazil relied upon an old school, 12 node, 240 core, CPU-only system using an improvised cardboard and packing foam rack. Now that’s impressive!
Just to show that Grid wasn’t a fluke, Team Brazil hammered the field again on the Nektarr++ app. Nektar++ can be used to simulate airflow over a Formula 1 car or to model how water flows through a corrugated pipe. It can also be used to provide theoretical estimates of dispersion and diffusion properties of discontinuous Galerkin schemes – which makes Nektar++ pretty handy. Galerkin schemes are cool.
CHPC took second place like a cheetah taking down a gazelle, and Tsinghua gets on the board with a third place finish. Honorable mention goes out to home team University of Hamburg for grabbing the fourth place slot.
The Mystery Application this year was Nek5000, a computational fluid dynamics solver that’s been in development for more than 30 years. You can use Nek5000 for fluid flows, combustion modeling, thermal convection, and a wide range of other things. Don’t try to use NEK4000, that’s the model number of a Microsoft Ergonomic Keyboard.
CHPC takes first place, like a helicopter gunship taking out a poacher. Tsinghua University was a close second with 1006 seconds elapsed time. Brazil’s UFPR makes another appearance, burning the rest of the field like gauchos cooking up a Picanha. A tip of the beer stein to Heidelberg for making another honorable mention for their fourth place finish on this app.
All of the applications we’ve seen so far account individually for 15% of a team’s overall score. The AI application, TensorFlow, is the big kahuna and is worth 25% of the overall tally. The TensorFlow challenge this year was fairly simple: have your machine recognize the greatest number of images per second with high accuracy.
We had a few teams go off the rails by making the challenge more complicated than it really was, resulting in disappointing scores. But that happens to some degree on every application, so it’s not unexpected.
ShanghaiTech pickled the field in wine, then stir-fried them into submission with their score of 3,244 images per second. Nanyang played the “buay steady” (check the Wikipedia Singlish dictionary for that one) to the rest of the competitors with their score of 3,157.8 images per second, while Tsinghua ran a close third with 3,081.6.
University of Poland makes their first appearance on the board with an honorable mention for fourth place, and since they were so close, we’re also giving home team Heidelberg some love for their score of 2,909.3 images per second.
The judges interview counts for 10% of the total overall score. Most teams did pretty well in the interview, given the average score of almost 8 and the median score of 8.3. During the interviews, judges ask the teams questions about how they optimized the applications, the steps they took to control power, how they selected their hardware, and how well they worked together.
CHPC tied with the UK’s own EPCC for top interview honors, while University of Hamburg snuck in for a second place finish. Tsinghua and UPC tied for third place, marking the first appearance for UPC on the charts.
So what have we learned? For one thing, you don’t have to win every category in order to win the event. Tsinghua didn’t win any of the applications or benchmarks outright, yet they took home the gold. Their win was based on a strong overall performance, with consistent second and third place finishes on each of the apps.
We’ve also learned that even if you don’t have the strongest hardware in the room, you still have a chance to earn cluster competition glory – just look at Team Brazil for proof of that. They had one of the least powerful clusters in the competition, yet they took first place on two applications and third on another. Who knows what might have happened if those kids would have had just a few GPUs so they could handle the TensorFlow app?
The ISC18 Awards Ceremony was the place to be on Wednesday afternoon. During the ceremony, the teams were introduced, winners announced, and trophies ladled out. I had a video camera and an audio hook up in the back of the room, so I was able to capture all the action. Gilad Shainer of the HPC AI Advisory Council was the master of ceremonies and I handled some of the color commentary and sorted through the various trophies. A grand time was had by all.
It’s been a great ISC18 Student Cluster Competition. The logistics were fantastic; as usual, each team got their equipment, all of it worked, and the competition tasks were handled perfectly. I’ve seen quite a few of these competitions and this one was one of the smoothest in history.
Next up we have the SC18 Student Cluster Competition in Dallas, Texas. It kicks off in mid-November, but you’ll be hearing from me well before then as I profile the competition, the competitors, and the tasks they’ll be facing. As always, stay tuned….