Mar 14

Print this Post

Let’s hack a ton! Eurohack Day 1

On March 6, ten teams met at JSC for the Eurohack 2017 to boost their applications on GPUs or port them to use GPUs at all. We had on average of 4 team members and 2 mentors per team. So we were topping around 70 hackers in one room. This year’s hackathon (see last years starting post) was again supported by the industry as Nvidia and IBM staff were around to help as mentors.

The Eurohack 2017 took place in the rotunda, second floor of the round building you see there.

After getting acquainted with the team, work started immediately for all Dresden mentors (7 in total).
In my team, we separated from the rest to dive into a code review and had the main developer explain to us the method of the asynchronator library and it’s advantages and disadvantages. As my team didn’t have any code running on the GPU yet, a lot of constructive discussion started immediately. As usual, a lot of arguments filled the room on each individual’s understanding of the computer hardware. At this point, this can be dangerous as it leads to premature optimization very rapidly. We agreed on a iterative approach guided by profiling and supported by unit tests (the latter being inplace already and even a benchmark suite running on a dedicated CI machine, very impressive).

The first day of the hackathon started off very smooth thanks to the wonderful organization by JSC. Dirk Pleiter, Andreas Herten, Michael Knobloch and others put the local infrastructure in place. Getting a training account on the GPU-equipped clusters, Jureca and Juron, was a matter of minutes. So everyone was ready to start! Interesting enough, I discovered that the entire environment modules infrastructure on Jureca is based on lua scripts (which I never saw before) that are auto-generated by a tool called easy_build. On Juron, all module files are written in tcl/tk as usual.

Just before lunch, Julian Miller invited all of us to participate in collecting data in a study to explore how much development effort is needed for porting applications to GPUs and other parallel architectures. After that Michael Knobloch gave us an introduction to profiling with score-p e.g. for use on the local clusters.

In the afternoon, Fernanda Foertner convened everyone to have the daily scrum. Every team was allowed to have 5 minutes to explain what they are after during the hackathon. The range of topics was very diverse (see the slides collection) this year. There were three Lattice QCD projects, three projects from brain science, fluid dynamics, life sciences, asynchronous mathematical methods and more. It’s interesting to note that 7 out of 10 projects already had accelerated code paths in production and some of them were already running on supercomputers. For these, the goal of the workshop was to squeeze more performance out of their application or to reach better scalability to run on large HPC machines. As simple arithmetic tells us, the remaining 3 projects were CPU-only code bases and needed to make the jump onto the accelerator this week. As I know from my personal experience, that’s a tough goal to reach.

In the evening, we spent a wonderful dinner at “Am Hexenturm” and finished the day with good food and drinks.

The evening dinner took place in a restaurant next this wonderful 14th century gate guarded by 2 defense towers called “Hexenturm” (english: Witchtower).

This blog marks one of the first in a small series that will cover our observations from this week-long hackathon. The blogs will cover suggestions straight from the lips of Nvidia engineers and best practises of seasoned CUDA developers that came up due to the issues the project teams had. If you think, I missed something or didn’t convey anything in the correct manner, please use the comment section below to indicate this.

Permanent link to this article: https://gcoe-dresden.de/lets-hack-a-ton-eurohack-day-1/

1 ping

  1. GPU Hackathon at Jülich Supercomputing Centre

    […] heavily involved into the Hackthon. They came over with seven mentors! Thanks for that! And they blogged about the event as well! Go and read their posts for some deeper insights into the challenges of GPU […]

Comments have been disabled.