Final cluster configurations have been set, and competitors in the ASC19 Student Supercomputer Challenge have started running the various AI models and HPC benchmarks that will determine who is declared champion. But if the first phase of the competition is any indication, there won’t be much rest for the next two days.
Students were forced to overcome numerous minor issues during the first two days of the 2019 ASC Student Supercomputer Challenge taking place this week in Dalian, China. It was an exercise in patience and perseverance for the teams, a good percentage of which are competing formally for the very first time.
Computers, of course, don’t always behave as they should. The key to dealing with digital adversity is to keep calm. Above all, don’t panic.
“There have been so many [things that are not expected],” said the advisor for one of the Chinese teams, which experienced difficulty in connecting a computer monitor to the cluster. “There are constant problems in the environment, so we are working through them.”
Multiple teams have run into unexpected errors with loading of the Computer Earth Simulation Model (CESM) software, which is one of the key tests that’s expected to test the resilience of the computer science majors that have gathered here at the Dalian University of Technology for the five-day event.
“The server environment [back home] doesn’t match the server here,” said one team member. “And the data doesn’t match, so now they have to try to new strategies for the competition.”
But first, it helps to power up the cluster. “They forgot to plug the computer in yesterday,” he said.
As the clock ticked down to the start of competition, most teams were busy testing their cluster against the various applications to determine their final setup. But one team was still doing the initial load of the environment after assembling the hardware, after discovering incompatibilities between the SSD drives they brought with them and the servers that are supplied by the Asia Supercomputer Community (ASC) and Inspur.
“We had many problems last night,” a team member told HPCwire. “This machine doesn’t recognize the disk drives. So we had to install everything from scratch. We are still installing some packages.”
After building their clusters, teams spent the first two days testing the applications and dialing in their server configurations. While there are caps on the amount of RAM and electricity they can bring to the game, competitors are free to use as many Inspur servers as they want. They can also bring any accelerators from the outside; most teams have some number of Nvidia GPUs.
Finding the optimal number of server nodes, as well as balancing the CPUs and GPUs, is the name of the game in the initial phase. Some of the applications — such as the Single-Image Super-Resolution (SISR) PyTorch-based deep learning model — will probably gobble up as many GPUs resources as you can throw at it, while CESM is a more CPU-oriented workload, multiple team members have said.
Just the same, just getting CESM running has been a struggle for some. “The first time we loaded CESM, we encountered a compiling problem, a compiling error,” said one team member. “The first two letters of CESM stands for compiling error. It’s very funny.” Yes, it is.
Another team discovered — to their great horror — that there was no Internet access from their Inspur servers. “We didn’t expect that the machines couldn’t directly connect to the Internet,” a team member said. “That was an unexpected challenge. So we had to make the bridge from one of our computers to that cluster.”
As the event rolls on, the competition will heat up — within the cozy confines of the 3,000 watt power limit, anyway. Some teams are battle-hardened veterans of supercomputer competitions, at ASC — last year’s event heated up so much the servers nearly melted down — as well as competitions at the Supercomputing Conference (SC) and the International Supercomputing Conference (ISC). But many are getting their first taste of international competition.
“Our situation is not good,” one team member said. “This is our first time in the competition. We don’t have a lot of experience. We are learning a lot about the rules and the applications. We are just trying our best.”
The ASC19 teams have been preparing for this competition for months. They’ve tweaked and tuned their models to run on their own servers in anticipation of the finals. The big wildcard is the actual data that they will be asked to process, which has been kept a secret until today.
“We try to prepare as much as possible,” a student competitor said. “As you know, we haven’t seen the final data. So we have to prepare a lot of pre-processing, post-processing, try different models and try to build the models as accurate as possible, then when we see the data we try to fine tune then to say which one is better, and we will use that. That’s the strategy we have.”
Some teams are anticipating that the actual data used for the CESM test will throw some wrenches into the gears — particularly if they take a GPU-heavy approach. Many teams are expected to idle their GPUs for these tests, although they will still be eating into their 3,000-watt power budget.
“The main problem is the data [for the CESM test] is very, very complex,” one team member said. “So if we want to transfer it to GPU, we have to do a lot of work. And the time consumed on the data transfer … is much longer than the computation.”
“It would take me at least a month or two month to finish the modification” to get it to run on GPU, his teammate said. “The time is not enough, so we cannot do too much…. We are majoring in computer science, not in climate science.”
The teams with experienced members have an advantage, as they can coach their less-experienced colleagues on what to expect during the finals.
“I think they are becoming more and more experienced,” one student veteran of the ASC Finals said. “They have learned a lot. I think they grew a lot for this test. I hope they get some great result out of the contest.”
The student teams that win the competition will no doubt get an academic boost — not to mention a monetary prize. But no matter who wins, one thing is for certain: all of the competitors will gain valuable experience in dealing with unexpected adversity that inevitably crops up in the real world.