Glasgow DPM Performance Tests
User:Jamie ferguson tested different kernels and filesystems on a RAID 5 box using a Dell Megaraid controller.
Tests
caveat - with the CERN kernel and the Vanilla kernel the number of gridftp streams were varied using the -T option in the glite-transfer-channel-set command, which, at the time of testing was broken, hence all transfers had T=4.
with the CERN2.6 kernel the number of gridftp streams were varied using the -g -p option of the file transfer script.
Summary
| ext2 | ext3 | xfs | jfs | |
|---|---|---|---|---|
| SL3.05 (Vanilla 2.4) | 170, 3.44 | 199, 0 | 291, 3.44 | |
| SL3.05 (CERN 2.4) | 217, 3.33 | 215, 0 | 316, 0 | |
| SLC3.06 (CERN 2.6) | 123, 0.78 | 164, 0.78 | 351, 0 | 285, 0.67 |
The above table quite clearly shows that xfs produces the highest transfer rate and has the added bonus of zero failure rate for both the kernels it was tested with! If you are considering migrating to an xfs filesystem then there are a couple of howto's located at Performance_and_Tuning which gives advice on how to change kernel and how to safely change the filesystem that data is located on.
Note that this was testing under severe load and that the transfer failures seem to be caused by excessive load on the server causing the setDone call to fail (i.e., the SRM daemons can't get enough CPU and i/o at the crucial time). It is a general good practice, if you can, to have no DPM filesystems on the DPM head node, but have that exclusively for the DPM daemons.
Results
The results of the above tests seem to indicate that to maximise the write speed onto disks sites should configure there filesystems to be xfs. For sites wanting to convert see the XFS Kernel Howto.
