Skip to content

Latest commit

 

History

History
152 lines (107 loc) · 4.98 KB

README.md

File metadata and controls

152 lines (107 loc) · 4.98 KB
Raw

GPU Programming with OpenACC

22-23 October, 2019

Course homepage.

Schedule

Day 1
9:00 - 9:30 Introduction to accelerators
9:30 - 9:35 (break)
9:35 - 10:30 Introduction to OpenACC
10:30 - 16:00 (self-study) exercises: hello world, vector sum, double loop
16:00 - 17:00 Q&A session
Day 2
9:00 - 9:45 Profiling and performance optimisation
9:45 - 10:00 (break)
10:00 - 11:00 Data management
11:00 - 16:00 (self-study) exercises: jacobi, heat equation (profile)
15:30 - 16:00 Advanced topic: Multiple GPUs with MPI
16:00 - 17:00 Q&A session

Lecture materials

The lecture materials of the course are available in this repository in the docs/ directory. Pre-compiled PDF versions of the slides are also available at: https://kannu.csc.fi/s/kYPBWXDd94PasdQ

Other useful material:

Exercises

Exercise materials are located in two different directories within this repository. In the exercises/ directory is a set of exercises prepared by CSC, while the exercises in the nvidia-labs/ directory are courtesy of NVIDIA.

Each directory contains a README.md file describing the exercise as well as any skeleton codes or other files needed for the exercise. Exercises for each topic are outlined in exercises/README.md.

For the exercises, you may use CSC's Puhti system with its GPU partition of NVIDIA V100 GPUs, or your own local system if you have suitable GPUs and an OpenACC compiler installed.

Downloading the exercises

To get the exercises you can clone the repository

git clone https://github.com/csc-training/openacc.git

This command will create a folder called openacc where all the materials are located. If the repository is updated during the course you can update your local copy of it with a git pull command.

Compiling with PGI compiler

PGI compiler commands are

  • pgcc for C
  • pgfortran for fortran
  • pgc++ for C++

OpenACC compilation can be enabled with option -acc. Note that without this flag, a non-accelerated version will be compiled.

In addition, on some systems you may have to specify the type of the accelerator e.g. with -ta=tesla flag. On Puhti, the default is tesla. If you run these exercises on some other system, you have to modify the type of the accelerator accordingly. You can check the type of the accelerator and recommended flags with command pgaccelinfo. If you want to get detailed information on OpenACC code generation, you can use option -Minfo=accel.

Running on Puhti

You can log into Puhti using the ssh command ssh -Y [email protected] where XXX is the number of your training account. You can also use your own CSC account if you have one.

Before compiling the exercises you have to load the correct environment module pgi as well as reload the git module to access this repository

module purge
module load pgi/19.7
module load cuda/10.1.168
module load git

Serial jobs can be run interactively with the srun command, for example

srun -n1 -p gputest -t 00:05:00 --gres=gpu:v100:1 --account=YYY ./my_program

If you want to run a longer job, you may also use the normal GPU partition, instead of the GPU test partition:

srun -n1 -p gpu -t 00:30:00 --gres=gpu:v100:1 --account=YYY ./my_program

If you are using CSC training accounts you should use the following project as your account: --account=project_2000745. To see what scratch disk space you have available, please run csc-workspaces.

Reservation

This course has also a resource reservation that can be used for the exercises. To use these dedicated resources, you may run your job with the --reservation=openacc2020 flag, such as

srun --reservation=openacc2020 -n1 -p gpu --gres=gpu:v100:1 --account=YYY ./my_program

Please note that the normal GPU partition (-p gpu) needs to be used with the reservation.

Running the visual profiler

The visual profiler is a GUI program which means we need to get the X session out of the compute nodes. The way this is done now is to first allocate a GPU for your use, then start an ssh session into the node you allocated. Once logged into the node you need to reload the modules and then you can run the visual profiler nvvp

salloc -N1 -n 1 --account=YYY -p gpu --gres=gpu:v100:1
ssh -X $(srun hostname)
module load pgi cuda
nvvp