Performance Basics

4.3 IO Configuration

  • Data on local batch T: disks are temporary. For batch jobs running under Cluster Controller, the system deletes "old" data periodically from the local T: drive. The best way to use this space is to have your job write on local disk while it is running, then copy the output files to your own directory on H: for permanent storage.

  • Latencies, i.e., the time to deliver the first word of data, vary greatly for each medium, but are much greater than any delays in the memory hierarchy.

  • Performance consideration: Since I/O is slow, it may be better to recompute a value than to store it on disk.
For more on file storage options and the advantages and drawbacks of each, see the talk on Data Management.