I/O tips

To most effectively make use of the disk resources within the cluster, consider the following scenario:

• A user edits, compiles, and tests a computational program interactively on amdahl.
• 30 GB of data sets are created (in 30 1 GB files), and uploaded into /project/groupname.
• A PBS script is created, with steps like the following:

  # Create scratch directory on the execution node
  mkdir /scratch/$PBS_JOBID
  # Copy data files into scratch directory
  cp /project/user/datafiles/file1.dat /scratch/$PBS_JOBID
  # cd into the /scratch directory
  cd /scratch/$PBS_JOBID
  # Run your executable, with input and output in the /scratch directory.
  $HOME/bin/program < file1.dat > file1.out
  # Copy files back home and cleanup
  cp file1.dat /project/user/output && \
  cd .. && rm -rf /scratch/$PBS_JOBID
  

• Taking steps like those above can help I/O-bound jobs, by not forcing every file read and write to do so via NFS, instead utilizing the compute nodes' locally attached storage.
• Taking steps like this will also reduce the possibility of a network hang causing your computation to terminate!