Scripts to generate Debian-based CUDA-enabled containers linked together with the SLURM job manager. Tutorial given at GPU Technology Conference 2014. See the slides for a detailed explanation. Markdown source for the slides is included.
Note that these scripts do not constitute a production configuration, but rather serve as a reference for building small scale CUDA clusters and running CUDA applications within LXC containers.
- Firewalled test machine
- CUDA hardware
- CUDA installed at
/usr/local/cuda
- Access as root
As root, on a preferably firewalled machine ,
./host_install.sh
./host_ready.sh
./install_cuda_node 0
./install_cuda_node 1
./install_cuda_node 2
./install_cuda_node 3
(cd /usr/local/cuda/samples/0_Simple/matrixMul && make)
ssh -i id_rsa [email protected] # or use password reported by install_cuda_node
srun /usr/local/cuda/samples/0_Simple/matrixMul/matrixMul
If during install_cuda_node
, ssh requests a password, ensure that the cloned
repository directory is owned by root and has permissions 700.