The ASC competitions are known for the punishingly difficult application tasks they ladle out to students at their annual student cluster competitions. As one long time student participant commented “The applications at ASC just keep getting harder and harder, it never stops.”
This year, the competition, held in Nanchang, China, is even more difficult since it’s a day shorter. The students only get two full days to handle their cluster business, rather than the traditional three. So what did the students do over those two days? Glad you asked. Let’s take a look their cluster competition tasks…
Teams have two days to put their clusters together both physically and mentally. They’re given nodes in boxes and bare racks and it’s their task to put the parts together and wire ‘em up. Then they have to load their software stack (o/s, libraries, management tools, etc.). The students can run anything they want, but the vast majority use CentOS as their operating system of choice.
The rest of Saturday and Sunday are consumed by testing and making final tweaks or wholesale changes to their cluster. Many of the teams at ASC18 overprovisioned their cluster by a wide margin, starting out with 10 or 12 nodes, even though they knew they were probably going to run fewer nodes when the competition started. Their reasoning is that they want to have the maximum amount of flexibility with their config and didn’t want to run the risk that they might not come up with a configuration that couldn’t cut the computational mustard.
A few teams configured systems that were too small to be competitive. They were, assumedly, looking at the label power on the nodes and trying to make sure they were right at the 3,000 watt mark. The better strategy is to over configure and then lower the power draw for each component in order to come in under the power cap. This means throttling back energy usage for CPUs, GPUs, fans, etc., and then testing their optimized applications under those conditions. The top tier teams have mastered these skills.
The competition started bright and early at 8:00 am on Monday. The first tasks were familiar ones: the HPL and HPCG benchmarks. There is a cash prize of 10,000 Yuan (about $1,500 US) for the team that wins the Highest LINPACK award, which raises the stakes even higher. HPCG is the more difficult benchmark to run and optimize, and it is also much more of a real world test than HPL, which makes it an important exercise in my mind. There isn’t an award for the highest HPCG, but I think there should be and will lobby that issue with all of the major competitions.
Teams also got the data sets for the RELION application. RELION is a life sciences 3D application that is used in the field of structural biology. It’s used to reconstruct biological structures in situations where x-ray, crystallography, and other methods fail to show the entire structure.
Monday also saw the unveiling of the ‘Mystery Application’, which is an secret application that the teams can’t prepare for. At ASC18, the mystery app was Siesta, which helps researchers perform efficient electronic structure calculations for molecular dynamics simulations. It scales linearly and is highly accurate.
Results for the benchmarks, RELION, and Siesta had to be turned in during the day on Monday.
On this, the last day of the competition, students were tasked with running two applications. The first was CFL3D, a computation fluid dynamics program that was developed by NASA Langley in the 1980s. It can be used for simulating and solving fluid-like flows in both 2D and 3D. It’s primarily used for external aerodynamics analysis.
The final application for the students is a challenge dubbed “Answer Prediction for Search Query”, which asks them to build an AI model that can understand a written document and correctly answer a question about it. Their model needs to be able to understand the context of the question in light of the document content – which is a highly sophisticated test.
The results for these applications had to be turned in before the end of day on Tuesday.
What happened on Wednesday is probably the most nerve racking part of the competition: the judge’s interview. This is where the student teams have to present their methodology and their competition results to a panel of HPC expert judges. After their presentations, the judges question the teams about their approach to the competition, why they did what they did, what they could have done differently, etc. Teams have to be very prepared prior to venturing into this lion’s den, as the judges aren’t known for cutting anyone any slack.
While the judges are all smiling in this picture, don’t let that fool you, they can be savage during the question and answer period with the students. They aren’t mean spirited in any way, but they have a knack for asking exactly the right questions to pinpoint a topic that the students aren’t quite prepared to discuss, or don’t want to discuss. It’s a gift they’ve developed through long experience with students and in the industry.
In our next article, we’re going to look at the results from LINPACK, stay tuned…