Cluster-wrestling kids pimp their HPC rides in Frankfurt

This one goes to 11: More rig pr0nz from students at ISC17

By Dan Olds, OrionX

Posted in HPC, 19th June 2017 10:54 GMT

HPC Blog Eleven teams at the ISC17 Student Cluster Competition will go head to head live in glamorous Frankfurt, Germany, this week. Yep, this one goes to 11.

For those few of you who aren't slavishly following these contests, here's a quick explanation.

A student cluster competition is where teams of undergrads design their own clusters, work with vendors to make their dreams come true, and then bring their machines to HPC events (SC, ISC and ASC) to compete live against each other on a series of benchmarks and real HPC applications.

The only constraint is that they are not allowed to use more than 3,000 watts – it's the law.

The organising committee has turned up the competition pressure to 11, as you'll see by the applications the students need to complete below.

Benchmarks: Nothing on this front has changed from previous competitions. They'll still be running HPCC (The HPC Challenge) benchmarks, with a special run of HPL (LINPACK) that will decide the Highest LINPACK award. They'll also be running HPCG (Conjugant Gradient), which will severely test their systems' ability to work with sparse matrix math and stuff. As benchmark designer Jack Dongarra once said to me: "HPL and HPCG are bookmarks – with HPL showing you the best your system can perform and HPCG showing you the worst performance you'll probably ever see."

Bring on the Science!

On the application side of the competition, the organisers are throwing curve balls and nasty sliders. Here's what the students will be swinging at...

FEniCS: If you need to run some partial differential equations, chances are you're a FEniCS fan. It's a quick way to translate scientific models into efficient element code. It features Python and CII interfaces, which makes it easy to get started, but it has a whole lot of power under the hood, which allows experienced programmers to make it scream.

Coding Challenge MiniDFT: First, MiniDFT is a plane-wave density function that's used for modeling materials. Give it a set of coordinates and pseudopotentials and it will compute self-consistent solutions of the Kohn-sham equations – which is sort of common knowledge, right?

The student teams ran a test case with an input set provided by the organisers before the competition. Then they have to run the same input case during the competition and prove that they can replicate a similar timing during the competition run. Then they can tune and optimise the code to run even faster, modifying any of the code they can, and then run the same input case.

TensorFlow: This is an open-source software library used for numerical computation using data flow graphs. Nodes in the graph represent mathematical operations while the graph edges represent the multidimensional data arrays (called tensors) communicated between them. This is one of the base programs developed by Google Brain engineers for the purpose of machine learning and deep neural network research. However, it's general enough to be used in a wide variety of other domains.

The HPC Advisory Council (big time sponsor and organizer of the ISC competitions) has been working with Baidu to design the deep learning tasks for the students. They're going to be running CAPTCHA image recognition using the TensorFlow framework with the Keras high-level neural network API. Teams will be scored on how well their model correctly recognizes CAPTCHA images from an untrained data set. The team with the highest degree of accuracy will get the most points.

Secret Application: This one is a secret. Even highly placed insiders refused to comment or speculate on the identity of the secret app. And this was after I plied them senseless with alcohol.

The scoring breakdown is 10 per cent for the benchmarks, 80 per cent for application runs, and 10 per cent for the interview with HPC experts on what the teams know and what they've learned.

There will be five awards given out for this competition. The Overall Championship goes to the team with the best performance on all tasks. Second and third places will also receive an award. Another award goes to the team with the highest LINPACK score, plus an award for the Fan Favourite, which is derived from online voting during the show.

As always, we'll be covering all aspects of this event like skin. Next up we'll meet the teams in touching and sensitive up-close-and-personal interviews.

Sign up to our NewsletterGet IT in your inbox daily


More from The Register

Huawei's 4-socket HPC blade server gruntbox gets Skylake mills

Beefier grunts from Chipzilla's latest and greatest

Want to know more about HPC apps? This explicit vid has some answers

HPC Blog Page through this profiler...

Stephen Hawking's boffin buds buy HPE HPC to ogle universe

But can COSMOS find a way to improve HPE profits? Hmmm

HPE gets carried array with HPC: Partners up with DDN

+Comment High-performance servers get data-pumping storage arrays

Teen Pennsylvania HPC storage pusher Panasas: Small files, fat nodes, sharp blades

Analysis Not your average Silicon Valley startup kid

HPE teases HPC punters with scalable gear

No you can't have full specs or pricing until next month

Love cloudy HPC? Microsoft does, slurps Cycle Computing

CPU-wrangler found fame on EC2, becomes Azure business

Ah, breathe that fresh alpine air. And look over there, a majestic HPC Advisory Council

HPC Blog Beastmode crew hosts three-day conference in Lugano, Switzerland

UK splurges £20m on six regional HPC centres

Why? For science!

ARM buys HPC software specialist Allinea to help devs code good

If Intel and IBM think HPC is a two-horse race, they need to think again