Here is a sample template for submitting your program to one of the GPU compute nodes:
#!/bin/bash -l # Name your job (used in the PBS output file names) #PBS -N GPUTEST # Specify the gpuq queue #PBS -q gpuq # Specify the number of gpus #PBS -l gpus=1 # Specify the gpu feature #PBS -l feature=gpu # Specify your resource account (use qr command to determine) #PBS -A Physics # Specify how much time you think the job will run #PBS -l walltime=00:15:00 # Have the job send you email when the job ends or aborts #PBS -M firstname.lastname@example.org #PBS -m ea # Join error and standard output into one file #PBS -j oe # Change to the directory that the job was submitted from cd $PBS_O_WORKDIR # Parse the PBS_GPUFILE to determine which GPU you have been assigned # and unset CUDA_VISIBLE_DEVICES gpuNum=`cat $PBS_GPUFILE | sed -e 's/.*-gpu//g'` unset CUDA_VISIBLE_DEVICES # if using PyCUDA set the CUDA_DEVICE environment variable export CUDA_DEVICE=$gpuNum # Pass the GPU number as an argument to your program program_name $gpuNum exit 0
Your program needs to know which GPU it has been assigned and the submission template above parses the contents of the file given in the envrironment variable PBS_GPUFILE to determine which GPU number your job has been assigned. You should pass the GPU number to your program as a command line argument and then set the default GPU in your code. See this page on command line arguments for examples on how to read a command line argument in C/C++,Fortran and matlab. If you are using PyCUDA, set the environment variable CUDA_DEVICE to the GPU number as shown in the job submission template above.