By Dan Olds
The Asian Student Cluster competitions held by the Asian Supercomputer Community (ASC) have always had a reputation of being particularly demanding for the student participants. However, for this year’s competition they’ve turned the difficulty meter to 11.
Maybe it’s because they weren’t able to hold the competition last year. Or maybe it’s because they have a record number of teams this year and didn’t want to make it easy. Or they could just have been in one of those moods when they selected the applications and data sets. Regardless, this is probably the toughest cluster competition I’ve seen in more than a decade of observation.
In addition to the competition standard HPL (LINPACK) and HPCG benchmarks, the ASC committee set up an ultra-fiendish slate of three real-world applications. Let’s take a look…
Language Exam: This is a natural language processing challenge that requires student teams to build their own deep learning neural network to have their clusters pass various English comprehension tests. The tests include standard exams for entrance into Chinese high schools, colleges, and for MA/MS candidates. In the training set alone, there are 4,603 passages of various length with 83,395 questions that must be answered. The test set promises to be even more challenging. The winning team will achieve the highest accuracy on the tests. Use of GPUs is highly encouraged.
The QuEST Challenge: This is a simulation of generic quantum computing circuits of multiple qubits in both mixed and pure states. For the preliminary competition, student teams simulated quantum circuits of 30 qubits using QuEST. This size of circuit coupled with at least 16GB of RAM, according to the ASC provided materials, is enough to crack the RSA encryption algorithm. Yikes. Students need to be careful if they’re using GPU acceleration – the answers might not be correct – so they’ll have to double check their output files to ensure accuracy.
The PRESTO Challenge: Ok, here’s where it gets weird. Radio pulsars are fast spinning stars that are sort of like lighthouses emitting strong beams of radiation. PRESTO (Pulsar Exploration and Search Toolkit) is a large suite of tools designed to discover and explore pulsars by looking at two dimensions: time and frequency. From a computer science standpoint, PRESTO is like a seafood chowder of code with something for everyone. It has FORTRAN, Python, ANSI C, and probably some Java, BASIC and Smalltalk thrown in for good measure. One of the students I interviewed said that the last revision of the FORTRAN code he saw in the application was older than he was. Ouch. It’s recommended that the students run PRESTO on CPU instances only, probably because there isn’t a reliable accelerated instance out there.
When I interviewed the students, they were split on which challenge was the most difficult. They mainly thought that QuEST and PRESTO were the toughest, but the Language Exam had it’s share of votes as well. There was one outlier team that figured HPL/HPCG were going to be hardest, but they might have been kidding around.
Dan Olds is chief research officer at HPC industry analyst firm Intersect360 Research and a student cluster competition enthusiast.