On Wed, 26 Oct 2005 at 9:27am, John Summerfied wrote
Tom Sightler wrote:write read ----- ---- ext2 81 180 ext3 34 222 XFS 109 213This suggests to me you have a problem with the journal. Placing the journal on its own disk seems good to me.
I placed the journal on a RAID1 of 2 10K RPM WD Raptor drives, and tested again -- it made little to no difference. ext3 write speed was about 37 MB/s. These servers need to go back into production, so it's time for me to stop futzing with this. I'm just going to do an all-software RAID50. I was going to use RAID6 with one big array, but hit a bad (as in, FS corrupting bad, see bugzilla #171354) resync bug.
I have a couple of questions, is access time important to you? If not, have you tried mounting with noatime? Also, have you tried increasing the size of the journal and the stride options (previous suggestion)? My guess is that your having journal contention and having an external journal might help. Another option, you could attempt to tweak the commit rate with the commit= mount option. The default is 5 seconds which seems too often for some disk arrays.
noatime made no difference, and neither did increasing the journal size or the commit interval. All my RAID50 testing was done matching the stride option to the chunksize of the software RAID0.
Thanks for all the suggestions folks have made along the way. -- Joshua Baker-LePain Department of Biomedical Engineering Duke University