Arvados

Feature #12630

Updated by Peter Amstutz over 3 years ago

Use proposal from In CWL: 

 https://dev.arvados.org/issues/17240 <pre> 
 $namespaces: 
   nvidia: https://www.nvidia.com/object/ 
 requirements: 
   - class: arv:HardwareRequirement 
     hardware: 
       - hwtype: nvidia:tesla-k80 
         hwcount: 1 
 </pre> 

 Implement: In container_request: 

 # requiring GPUs in <pre> 
 runtime_constraints: 
   hardware: 
     - driver: nvidia 
       type: tesla-k80 
       count: 1 
 </pre> 

 Scheduled as: 

 <pre> 
 sbatch --gres=gpu:tesla-k80:1 
 </pre> 

 Crunch-run, setting up container, when RuntimeConstraints.GPU.Driver == nvidia: 

 * Inject devices /dev/nvidiactl, /dev/nvidia-uvm and /dev/nvidia[0-999] into container record  
 # crunch-run interpreting the requirement and launching the docker and singularity containers with CUDA support 

 
 * Inject environment variable CUDA_VISIBLE_DEVICES 

Back