How do I use a /scratch at a local disk?

A: A local working directory /scratch is available at a local disk of each worker node, and is intended for temporary files. Upon a job start, the batch system creates a unique subdirectory /scratch/condor/dir_ID/, which is accessible via the $_CONDOR_SCRATCH_DIR variable. When the job finishes, the subdirectory is automatically deleted. This way, the local working directory /scratch is continuously cleaned.

Very intensive I/O operations during the job’s run can cause overload and slow response of the NFS server. We highly recommend to use the local /scratch to avoid these problems. The best practice is: read all input files from the NFS server only once when your job starts, use /scratch at a local disk of an assigned worker node as your working directory, and write all output files back to the NFS server when your job completes.

A default job’s limit for /scratch is set to 20GB per one core. You can appropriately adjust a default requirement in a job_submission_file, e.g.

request_disk            = 5G
request_disk            = 30G

In case a job exceeds a required value for scratch, the job will be on-hold for one day and then will be killed by the batch system. You can modify scratch requirements during the hold state and release the held job (see Hold State).