Nice cluster, kid. But can these supercomputing students actually predict the weather?
It's hard enough for the Met Office...
HPC blog Your correspondent wrote this blog during the action at the ISC’16 Student Cluster Competition arena, located on the trade show floor of the ISC’16 conference. So far, everything is going well for the teams – everyone has their hardware and finished their basic set-up over the weekend.
Yesterday was when the rubber met the road, when the talk stops and the computing begins.
At 3pm, the competition kicked off with a Le Mans-style start: each student team had to scour the show floor to gather team t-shirts hidden in various booths. Once the team has all of their shirts, they can begin the competition.
From then, they have until 8:20pm to optimize, run, and turn in their HPCC and HPL (LINPACK) runs. The HPCC benchmark makes up 10 per cent of the overall team score.
Since its inception, the ISC competition has gained a rep for high LINPACK scores, most recently setting cluster competition records in two of the last three years. The team with the highest HPL takes home the prestigious Highest LINPACK trophy.
Scientific apps and mystery challenges
The ISC competition is more a set of sprints rather than a marathon. Students have a set amount of time, usually a morning or afternoon, to optimize and turn in the results from each application. This places a premium on how well the students know each application and how to adjust it to achieve maximum performance.
In this competition, the students' scores on the various applications counts for 80 per cent of their overall score – so it’s critical that they complete each app and turn in a valid result. This year, the ISC organizers are serving up a particularly challenging slate of apps…
Let’s take a closer look…
- Graph500 is a very data intensive benchmark that will stress the hell out of the communication subsystem of a cluster. Graph problems are tough for traditional supercomputers to handle, since graph problems don’t use floating point or typical compute routines. A typical graph-like calculation is the old Traveling Salesman Problem, where you’re trying to find the most efficient route through a large set of locations. Each permutation of the route has to be tested against other alternatives, which takes a lot of compute power.
There’s a twist in the competition on this task: for the first time ever, students have to write their own algorithm to solve the graph problem. They aren’t allowed to use the reference implementation – they have to provide their own approach. Teams will be judged on the quality of their code as well as their solution to the problem. It’ll be interesting to see what happens on this application.
- Splotch is a ray tracer used to visualize SPH (Smoothed Particle Hydrodynamics) simulations. It is typically used in fluid simulations, but is also a great tool to visualize astrophysical phenomena. It’s a demanding application since it can handle massive datasets where every particle is ordered as a function of their proximity to other particles. In short, if you have a bunch of data about how a galaxy forms and you want to visualize it, get yourself some Splotch.
- WRF is a weather forecasting and atmospheric research model that we’ve seen in several other student cluster competitions. It can produce weather simulations using either real world data or input conditions. It’s well documented and scales to hundreds or thousands of nodes.
ISC was the first cluster competition to introduce the concept of the “mystery application” or “secret task” that isn’t revealed until the day of the competition. In the past, students have had entirely new applications thrown at them, or had to reconfigure their hardware, or faced a race to solve a dataset using the least amount of energy possible. The mystery task/app is still a mystery to me, it’ll be revealed on Tuesday or Wednesday.
In our next report, we’ll outline what hardware the teams are running, then we’ll roll out our video “Meet the Teams” features. Stay tuned….