GPU Computing

Here you’ll find notes on accelerating computing with graphics processing units. I attended a course given by Rainer Spurzem at Heidelberg University in August 2017. A good resource is „CUDA by example“ by Jason Sanders and Edward Kandrot. While CUDA (used in the course) is proprietary software from NVIDIA, there is the OpenSource alternative OpenCL. There’s also a nice free course on parallel computing with CUDA by NVIDIA on UDACITY with this repo on github, make sure to check it out :D

How to run CUDA on Kepler

Kepler is a Supercomputer with 12 Nodes, 2500 Cores each. There are two steps to get onto kepler:

1. Log onto the gateway cassiopeia: ssh -l username welcome.ari.uni-heidelberg.de
2. From there, log onto kepler: ssh -l username kepler

Logged into your account you can get your tasks done:

1. Save your CUDA Code in a code.cu file.
2. Load the right Cuda module module load cuda/7.5
3. Compile your code with nvcc -o code code.cu
4. Run the sbatch-script with sbatch gpu_script.sh

Note, that the gpu_script.sh has to include

• the right version of the cuda module
• the name of your compiled program