The plot below is a measure of the read and write rates a single user would experience via the PDSF batch system. Jobs are submitted sequentially every hour either to the debug queue or to the Mendel queue. If a jobs doesn't finish in 8 minutes, it is killed and a -1 rate is written out.

The read rates are calculated by copying a directory containing 2 files totaling 274 MB (for Mendel it is 4 files totaling 548 MB) from the eliza directories to the $TMPDIR on the node running the job. The write rates are calculated by untarring a tarball on the eliza directories. The write rates are typically around a factor of two slower than the read rates, because the data still has to travel to the compute node and then back to the eliza for writing. The I/O rates are taken from the ganglia monitoring and serve as a measure of the amount of user I/O going on at the time the test was done.

This is intended to give a feel for the typical rates a user would see and is not intended to show the maximum rates that the elizas can generate by initiating multiple parallel transfers, for instance.



Tell my wife I said hello.