Experimentation. Investigation. Collaboration. It was all on the table at the first GPU Hackathon in 2019, where over five days in March dozens of students, researchers and industry experts came together at Yonsei University in Seoul, South Korea, to push the limits of software and science using GPU acceleration.
The GPU Hackathon program first launched in 2015 to help scientists, researchers and students within computer, computational and data science domains to accelerate and optimize their applications and projects on GPUs. Since then, over 700 domain experts from high performance computing centers around the world have been trained and nearly 200 HPC applications using directives to port to GPUs have been tracked.
The Seoul hackathon was the first of over 15 planned events worldwide that will take place in countries such as Australia, Germany, India and the U.S. It was hosted by Yonsei University, the Korea Institute of Science and Technology Information (KISTI), Oak Ridge National Laboratory in the U.S., and NVIDIA.
Participating teams included students pursuing different majors and hailing from universities across Korea. They engaged in a weeklong design sprint, with computer programmers and software developers optimizing and accelerating the performance of their projects.
The participants used KAT, a GPU cluster system of NURION, KISTI’s fifth-generation supercomputer. Each team was assigned two mentors out of more than a dozen experts attending from KISTI, NVIDIA and PGI. Using the supercomputer and consulting with their mentors, the teams worked together to improve programs they’re already using or newly developing.
“This hackathon was a great opportunity to see the potential of my new program,” said Joon-goo Lee, who attends the graduate school of Seoul National University.
“I wasn’t sure how much this newly created program would help us, but what’s more important is the fact that we challenged it and gave it a try,” said Deokhoon Kim, a graduate student at the University of Ulsan.
The participants presented their optimized programming before an expert panel on the last day of the event. The panel then provided feedback on the direction the programs should head and further ways to enhance performance.
Team DCSLab, consisting of seven graduate students from Seoul National University and Sookmyung Women’s University, presented their program to calculate gravitational waveforms in space.
“We’ve observed 11 gravitational waveforms to date, but we’ve been developing a program to see more of them,” said presenter Joon-goo Lee. “Using deep learning, we want to build a program that helps us do iterations multiple times without a computational cost burden.”
This work caught the eye of one business mentor, who, on the spot, suggested a potential collaboration combining DCSLab’s in-depth knowledge and data with his company’s programming technology.
A team named FVCOMik unveiled a prediction model for sea status forecasting. “We are working on GPU parallelization to provide more accurate and faster forecasting,” said presenter Dong-hoon Kim from Inha University in Incheon. “This event was an excellent opportunity for us to learn about advanced technologies that we haven’t heard of and also to enhance the performance of our prediction model.”
Team Ulsan showcased a prediction model that figures out how much air and water resistance is generated when ships, airplanes and automobiles are in operation. During the hackathon, they sought out ways to cut the time to build the model from months to weeks.
A team of robot developers from LG Electronics, called ELSA, presented a deep learning-powered programming model that helps robots rapidly recognize objects in images taken by cameras.
“At this event, participants met some of the leading GPU experts at both home and abroad,” said Boo-young Ahn, head of the KISTI Science Data School Center. “They enhanced the performance of their codes by optimizing and parallelizing them. I hope these achievements will serve as a solid foundation for them to grow into experts in their fields of study.”
For more information on GPU Hackathons and to view future events, visit www.openacc.org/events.