Department Logo Information Services and Technology

Advantages of Using Local Scratch Space

Faster HPC - HPC performance boosts gained by using scratch space, and UCS' policy for deleting scratch space files.


You can greatly improve the speed of your HPC applications by making use of node-local scratch space instead of using your home directory for temporary files (assuming your HPC application uses temporary files). A /scratch directory is physically local on each HPC node specifically for this purpose. The local /scratch filesystem provides far faster file I/O than your home directory because the home filesystem is physically located on the HPC head node, and intra-node data transfer is far slower than local hardware data transfer. This advantage depends on your HPC application allowing you specify a specific directory for scratch or temporary files. Unlike input or output/results files, the application deletes the scratch file as part of normal operation. Scratch files should not be confused with checkpoint files. Checkpoint files are intended to be reused if your HPC application crashes, whereas scratch files cannot be used to resume applications. Use of /scratch space for checkpoint files is encouraged, but you must find and move your checkpoint files to your home directory on your own. We do not at this time provide an automated facility to do this. If you use Gaussian 03 for your research you can use your existing or GJF to use /scratch without any modification just by using our gauss_submit command instead of using qsub and g03 (or g03_qsub.sh). Authorized users of Dr. Bozzell's reserved nodes on Kong are automatically allocated additional access to /bscratch by gauss_submit. Please contact us if you want help adding automatic /scratch to a different application. Do not use /tmp for your scratch space. Our deployment has /tmp filesystems much smaller than /scratch and /tmp is always erased on system reboot.

Scratch File Deletion Policy

  1. Files more than 72 hours old are automatically deleted from all nodes' scratch spaces, subject to 3 below
  2. No notifications or backups are provided.
  3. If the scratch files' owner has any process running on the same node, those files are never deleted, even if :
    • the application runs longer than 72 hours
    • the application is not updating your local scratch files
    For bullet item 3 you need only be concerned with the scratch file deletion policy if your application crashes and you need the scratch files for diagnostic purposes, or if you are storing your checkpoint files in scratch space.