OpenACC in Umeå
Posted: 2 May 2013 | 14:08
I'm just back from giving my first PRACE training session as part of the joint PRACE/SNIC* Spring School based at Umeå Univeristat in northern Sweden. For those whose Nordic geography is like mine, here's a map...
This school comprised three different sets of speakers from PRACE Advanced Training Centres (PATCs) giving both lectures and hands-on sessions in new and emerging technologies for accelerating applications.
Days one and two were run jointly by Intel and CSC and were based around the Intel Xeon Phi. The Xeon Phi is a new type of offload device from Intel and whilst it may appear similar to a GPU in the frame of HPC, it is quite different, being built entirely for doing work. With a good mix of hands-on sessions using the hybrid system at CSC and lectures from Michael, Mykka and Olli-Pekka we learned about how to leverage performance from this device. As an added bonus, Michael sits on the OpenMP board with EPCC's very own Mark Bull and gave us an insiders rundown of the new OpenMP 4 standard.
On day three, it was my turn! I spent some time last year looking at how to program accelerators (such as GPUs) using the various directives-based approaches and have followed, and been involved with, OpenACC for some time. So, this was my topic. Somewhat terrifying, but good fun, I guided the 30 or so participants (ranging from Computational Scientists to HPC programmers and admins) from the basics of adding a few directives to an existing application to performance tuning.
Finally, on day four the speakers from BSC focussed on their ARM-based system, Tibidabo. This was quite different to most of the HPC machines we had used, being very power efficient. Power efficiency is a huge topic when considering the exascale and it was good to see how much performance one can squeeze from a small machine.
Left to right: me, Mikko, Bö, Michael, Olli-Pekka & Jerry
As you can see, it was quite warm during the week (around 10C) although I was told there was 2 feet of snow not long before our arrival!
The benefit of training sessions such as these is twofold: teaching those who use and who help others use HPC systems, and getting to meet others from different HPC centres. Networking is really important as only by meeting users can we understand what problems they face and how best to help them.
My thanks must go to Bö Kagstrom, Mikael Rännar, Jerry Eriksson and their colleagues at HPC2N for hosting this school and introducing me to reindeer steaks.
See also Ruyman Reyes Castro's OpenACC blog post.
To read about Xeon Phi at EPCC, see Iain Bethune's blog post.