GPU node
Queue Classes
[] GPU nodes use the following queue classes.
class | no. of nodes | no. of chuncks | GPU | execution duration | no. of concurrent jobs | no. of concurrent jobs/user | no. of submitted jobs/user | priority |
---|---|---|---|---|---|---|---|---|
G-SINGLE | 1node | 1 | ~2 | 2day | - | - | - | 1 |
G-MEDIUM | 2node | 1~2 | ~4 | 1day | 1 | 1 | - | 2 |
G-LARGE | 4node | 1~4 | ~8 | 12hour | 1 | 1 | - | 3 |
Interactive Jobs
[] Interactive jobs can be generated using the batch scheduler.
% qsub -I -q G-SINGLE
After logging in to the GPU node, please set the environment using the module command.
Sample execution
[] 0. If the workspace directory doesn't exist, please make the directory.
% mkdir /work/<UserName>
1. CUDA sample,copy the job-script
% cp -r /opt/cuda/6.5/samples/ /work/<UserName>/cuda_sample
% cp /work/Samples/PBS_scripts/pbs_gpu.sh /work/<UserName>
2. Running an interactive job,compiling on the GPU node
% qsub -I -q G-SINGLE
% cd /work/<UserName>/cuda-sample/1_Utilities/devicequery
% module load cuda
% make
3. Exiting from the interactive job
% exit
4. Submitting the job
% cd /work/<UserName>/cuda-sample/1_Utilities/devicequery
% cp /work/<UserName>/pbs_gpu.sh .
% qsub pbs_gpu.sh