The 4th annual HPCAC-ISC Student Cluster Competition (SCC) was a joint event hosted by the HPC Advisory Council (HPCAC) and the organizers of the International Supercomputing Conference (ISC). The SCC event was held during the 2015 ISC High Performance Conference (ISC 2015) in Frankfurt, Germany from July 13 to 15, 2015.
The competition featured 11 small teams of undergraduate students from nine different countries that compete to showcase the small clusters of their own design on the ISC’15 exhibit floor. The participating teams for this ISC 2015 Student Cluster Competition were:
- Centre for High Performance Computing (CHPC), South Africa
- A hybrid team of Purdue University, USA and Universidad EAFIT, Colombia
- Chemnitz University of Technology, Germany
- University of Science and Technology of China (USTC), China
- University of São Paulo (USP), Brazil
- University of Hamburg, Germany
- Jamia Millia Islamia University (JMI), India
- University of Tartu, Estonia
- Universitat Politecnica de Catalunya (UPC), Spain
- Tsinghua University, China
- National University of Defense Technology (NUDT), China
In this three-day event, the university teams raced to demonstrate the greatest performance across a range of HPC benchmarks and applications under the 3000-watt power budget. The teams were given the HPCC and LINPACK benchmarks on the first day of the competition, while the rest of the other two days focused on the HPC application performance.
• CONFERENCE AGENDA ANNOUNCED: The highly-anticipated educational tracks for the 2015 R&D 100 Awards & Technology Conference feature 28 sessions, plus keynote speakers Dean Kamen and Oak Ridge National Laboratory Director Thom Mason. Learn more.
The applications selected for this year’s competition complement very well with the GPU and the accelerators. PyFR is developed by Imperial College of London — it is an open-source Python based CFD application for solving advection-diffusion type problems on streaming architectures using the Flux Reconstruction approach of Huynh. PyFR has supports for NVIDIA CUDA, Intel Xeon Phi, and OpenMP thread models with MPI. The second application selected was LAMMPS, which is a classical molecular dynamics code. The newly introduced KOKKOS package for LAMMPS has support for NVIDIA CUDA, Intel Xeon Phi as well as OpenMP thread model. OCTOPUS is a scientific program aimed at the ab initio virtual experiment which calculates quantum mechanics within density-functional theory (DFT) simulation; it also has support for accelerators as well.
Below are highlights some of the events that took place during this year’s competition:
- Tartu Overcame Missing Hardware
There is no competition without upsets and surprises. The team from University of Tartu of Estonia suffered from numerous missteps but was able to overcome them about two hours before the submission deadline on the first day. Tartu encountered some shipping logistic issues which caused them delay of their power supplies for their cluster and rack. It was discovered that about 18 power supplies were missing during the weekend before the competition, which prompted the team to go out and purchase power units on Monday (All stores are closed on Sunday in Germany). The Tartu team was also spotted cutting power cables and assembling the cable plug on their own to make it work with their servers. We applaud their efforts to recover from the situation in such a short time and not give up.
- Tight race on LINPACK between JMI and Chemnitz
The team from Jamia Millia Islamia of India ended up grabbing the Highest LINPACK award. That LINPACK score beat the previously set record of 10.1 TF established last year by the EPCC team. Close behind in the second highest LINPACK score was the team by Chemnitz with a LINPACK score of 10.31 TF.
- Battle of the hardware: UPC fired up 135-node ARM cluster, Tartu featured AMD APU, USP armed with both Xeon Phi and NVIDIA GPU, CPU-only cluster became minority
Alternative architectures are also being deployed against the mainstream Intel architecture. Universitat Politecnica de Catalunya (UPC) team of Spain featured ARM processors backed by the Mont-Blanc Project. University of Tartu came up with an AMD A8-7600 APU instead. Hybrid CPU-GPU clusters were featured on almost all teams. One exception is the University of Hamburg that featured a CPU-only cluster exclusively. In an effort to capture even more advantage by the accelerators, University of Sao Paulo of Brazil was the only team that used two Intel Xeon Phi and an NVIDIA GPU on the same machine of their cluster.
- The Secret Mission
The secret mission was a task given on the last day that gave teams an opportunity to reconsider the hardware configuration impact on performance by removing any hardware on their systems to give them a winning edge. The criteria to win this task is using the minimum power to complete the specific PyPR workload within a 45-minute time period. The challenge for this mission is for the teams to understand the characteristic of the workload which is unknown to them ahead of time.
Changing hardware configuration can be an unfamiliar territory for students. The decision to change hardware and to perform reconfiguration is done without assistance from their advisors or vendors. Team advisors can only look from afar at the student lounge, while power monitoring is shown as the secret mission takes place.
- Hybrid Team of Purdue and EAFIT won the Fan Favorite Award
The Fan Favorite Award was given to the team that received the most unique votes during the SCC. The voting accelerated towards the end, as many teams ramped up the voting efforts, but Purdue/EAFIT enjoyed a lead at the end. It appeared the hybrid team of Purdue and EAFIT ended up with surging votes from Facebook, and majority of those belong to the Colombian fans who made a huge impact in their total votes.
- Tsinghua led in Secret Application; grabbed Grand Champion title
Tsinghua is a very experienced team for student cluster competitions. Tsinghua and NUDT were the two ASC15 teams that won the tickets to enter the ISC SCC race. Tsinghua led on some applications but got the winning edge on the second day’s secret application. The secret application is an unknown HPC application, which the teams do not know about until the day of the competition. The secret application was the benchmark called Graph500, which appears to be no challenge for the Tsinghua team.
- Defending Champion CHPC conceded to the 2nd place
CHPC, the defending Champion of the HPCAC-ISC SCC from the last two years, ended up winning the second overall place.
- Close race for the 3rd place
The race for the third place was almost in a deadlock among three teams. We had to look closely and hard among those three teams to determine a winner. USTC came out winning but the other two teams put up a very good fight, which is worth some notable mentions. Chemnitz University of Technology got excellent scores for the ones they were able to run. Only three teams managed to score on all of the applications. Besides Tsinghua University, Hamburg and USTC were the only teams able to score in each and every input data given for all three days. Even though University of Hamburg managed to be very consistent for most of the application runs, USTC came out slightly better in overall scores. USTC ended up grabbing the third overall winner award.
Vendors and Developers Support
We greatly appreciate the support many of the hardware sponsors provided to the teams. We are glad to see that most teams have been able to receive some support for the hardware used at the competition.
We are also grateful that the developers of PyFR, LAMMPS and OCTOPUS contributed their time and efforts for the months leading to this event, which included providing guidance for students and troubleshooting problems and inquiries by students. In particular, we greatly appreciate the contributions by Peter Vincent and Freddie Witherden of the Imperial College of London for assisting students on PyFR and also making themselves available for the students during the competition. We also appreciate the help from Christian Trott of Sandia National Lab for the support and help on the LAMMPS-KOKKOS application, as well as Xavier Andrade of Lawrence Livermore National Lab for the support we have received on OCTOPUS.
The goal for the HPCAC-ISC Student Cluster Competition is to give a unique opportunity for the students to learn, experience and demonstrate how HPC can influence our world. The months of preparation leading to the competition helped them enhance their knowledge, which will become useful for their studies and professional careers. We would like to see more teams get involved in the competition for the next year.
Pak Lui is ISC Student Cluster Competition Co-chair, HPC Advisory Council.