AsianScientist (Jan. 31, 2018) – While most young people are digital natives by default, not many of them can say that they know how to build a supercomputer within 48 hours. A six-member team of undergraduates from Singapore’s Nanyang Technological University (NTU) has not only done that but gone one step further and won the hotly-contested Student Cluster Competition held at the annual SC17 conference in November 2017.
Now in its 11th year, the competition invites students to submit proposals for custom-built supercomputers that are constrained by a 3,000-watt power limit. Selected teams then travel to the conference site in Denver, set up their machines and put them through a series of tests designed to stress their supercomputers’ processing capabilities.
The NTU team, sponsored by the National Supercomputing Centre of Singapore, beat 15 other finalists to emerge champions. Not only did they achieve the best performance on both the traditional LINPACK benchmark and the newer High-Performance Conjugate Gradient (HPCG) measurement, they set new records for both benchmarks.
Asian Scientist Magazine speaks to team captain, Mr. Liu Siyuan; vice-captain, Mr. Shao Yiyang; and their professor-in-charge, Associate Professor Francis Lee to find out more about what went into the team’s record-breaking performance.
How did the team prepare for the student cluster competition?
Liu: We had weekly meetings to discuss the various high performance computing (HPC) technologies and how to utilize them for the competition. We also consulted domain experts in fields relevant to the applications used in the competition. To practice optimizing the applications we had, we tried them out on small training clusters as we didn’t have access to the machine used for the competition until a week before the event.
Part of your strategy was to turn off the fans on your devices for a while, could you explain why you did that?
Shao: The competition had a power budget of 3,000 watts and we wanted to reduce the power consumption to gain the maximum performance from our hardware. We figured that the fans would consume quite a large portion of the power, and that short applications would not generate much heat, so we tried to lower the fan speed when running certain applications.
The problem with that was if the temperature was not controlled properly, the system would shut down altogether. So we had to find a balance between maintaining the temperature of the hardware and its capacity.
Which component of the competition did you find the most challenging?
Liu: The Born application [a seismic imaging app used to identify oil and gas reserves] was by far the hardest. This application is usually used by oil and gas companies to find oil and gas underwater and is a very computationally-extensive application.
We first tried to run the program on central processing units (CPUs), and found that it took a few hours just to go through a few data sets—we had more than a thousand of those data sets to handle during the competition. Obviously, if we ran Born on just the CPUs, there was no chance of completing it within the set time.
Instead, we spent two months porting the application to graphics processing units (GPUs) and the speed-up was significant. After we ported Born to the GPUs, it took something like 10 minutes for each data set. The GPUs enabled us to finish everything, but we spent a lot of time making this port, and that’s why it was most challenging of all the applications.
Why was it so difficult to port the application from CPU to GPU?
Liu: The porting process requires you to revise all the code, changing it from one code architecture to another, and GPU programming is not an easy task. It’s also very hard to identify problems with GPUs, unlike normal CPU programs. So we spent a lot of time programming the code to make it work correctly. The other applications were mostly about our understanding of the application and didn’t require us to change the code.
Was the team expecting to win the first place?
Liu: Before the competition, we felt we had a chance of winning, because we had put in a lot of preparation beforehand and we did much better than last year. I was hoping to get good results, but it was still a little surprising when we won as overall champions.
Lee: The other top three teams are all very strong, including teams who had won the competition before. We knew we worked hard, and kept our fingers crossed. But to come out on top and winning like that was a real bonus.
How did you feel when you were announced the winner?
Liu: I was ecstatic! My first thought was that it showed that we have the ability to compete. The win validated the hard work we put into this, so that was very satisfying.
What was the biggest lesson you learnt from participating in this competition?
Liu: It was definitely a great opportunity for us to try out a lot of things. It is a very comprehensive competition that tests you on all aspects, and we learnt a lot of extra things that were not necessarily taught in school. That has helped us become more competitive in the future as well.
Shao: The competition was also a great time for us to try our hands at all the latest technologies. We also networked and met people including seniors and those who had put in so much effort to make the competition happen.
What advice do you have for future teams that are going to take part in the competition?
Liu: My advice would be to try to do your best in all aspects. You cannot win the competition by just doing well in one thing, you need to put in enough effort to prepare for all the aspects involved, such as the interview, the poster presentation and non-technical aspects as well.
Shao: I think it is most important to keep up the passion you have in HPC and computer science, and try to enjoy the process of preparation for the competition. While at the conference, take the opportunity to meet people and try to learn as many interesting things as you can along the way!
———
Copyright: Asian Scientist Magazine. Photo: SC Photography.
Disclaimer: This article does not necessarily reflect the views of AsianScientist or its staff.
