gpu-lecture-task-parallelism

Lecture given at TU Dresden on task parallelism with GPUs (CUDA, OpenACC).

Working on Taurus (if needed)

First, establish an interactive bash session on a gpu host (attention: the reservation flag will only work during class hours, omit them anytime else):

for Dec 16, 2019

$ srun --reservation p_lv_gpu1920_323 -A p_lv_gpu1920 -t 1:30:00 --mem=4000 --gres=gpu:1 --partition=gpu1-interactive --pty bash -l

for Dec 17, 2019

$ srun --reservation p_lv_gpu1920_324 -A p_lv_gpu1920 -t 1:30:00 --mem=4000 --gres=gpu:1 --partition=gpu1-interactive --pty bash -l

Second, please setup the correct environment (the defaul CUDA on taurus is version 10.1.243):

$ module add modenv/scs5

Module GCCcore/6.4.0, zlib/1.2.11-GCCcore-6.4.0, cURL/7.58.0-GCCcore-6.4.0, expat/2.2.5-GCCcore-6.4.0, XZ/5.2.3-GCCcore-6.4.0, libxml2/2.9.4-GCCcore-6.4.0, ncurses/6.0-GCCcore-6.4.0, gettext/0.19.8.1-GCCcore-6.4.0, Perl/5.26.1-GCCcore-6.4.0, git/2.18.0-GCCcore-6.4.0 unloaded.
Module GCCcore/6.4.0, zlib/1.2.11-GCCcore-6.4.0, cURL/7.58.0-GCCcore-6.4.0, expat/2.2.5-GCCcore-6.4.0, XZ/5.2.3-GCCcore-6.4.0, libxml2/2.9.4-GCCcore-6.4.0, ncurses/6.0-GCCcore-6.4.0, gettext/0.19.8.1-GCCcore-6.4.0, Perl/5.26.1-GCCcore-6.4.0, git/2.18.0-GCCcore-6.4.0 loaded.
$ module load CUDA
Module CUDA/10.1.243 loaded.

Working with the code

Note that the examples use C++11. please make sure that the installed host side compiler supports this standard. If your host uses gcc as the default c/c++ compiler, any version between 4.9 and 7 should be fine with CUDA 10, see here for compatibility matrix.

To prepare for class, change your working directory to where your code lives and checkout this repo :

$ cd /path/to/where/I/want/to/work
$ git clone https://github.com/psteinb/gpu-lecture-task-parallelism.git
$ cd gpu-lecture-task-parallelism

Let's start with the acceptance test:

$ cd 0_getting_started/
$ make
$ ./test_simple_increment

If this test passes, you are ready to go, if not, diagnose the problem. Feel free to post an issue to this repo if you are unable to solve the problem.

psteinb / gpu-lecture-task-parallelism Goto Github PK