Hi! Are these results i got from my benchmarking "normal" or is there something strange happening in my system? This box is going to be a mailserver, so i tested various fs to decide on which i should utilize. I personally dislike reiserfs because it has proven somehow unreliable on our servers (although it got better in the more recent kernels), ext3 would be quite good (fsck'ing partitions of 70 gb with a sh*tload of files is really painful) but has shown to be horrible slow on my setup. The test-partitions are on a ICP Vortex GDT8523RZ which runs in a RAID10 setup, the filesystems got created/mounted with standard settings (except 1024 byte blocksize on ext2/3 partitions). The noatime mount option hardly affected the performance, switching from ordered to writeback in ext3 also didn't change the results notable. I didn't dare to play around with reiserfs mount options because i hardly have any knowledge about their drawbacks. Kernel is 2.4.19-pre5-ac1, noticed no big differences to plain 2.4.18. I used settings which would reflect an average mailservers usage (using maildir format of course): 51200 files, ranging from 100 byte to ~10kbyte distributed over 1000 directories (-s0 -n50:10000:100:1000). Here are the bonnie++ results i got: ext2: Version 1.02a ------Sequential Create------ --------Random Create-------- testMX -Create-- --Read--- -Delete-- -Create-- --Read--- -Delete-- files:max:min /sec %CP /sec %CP /sec %CP /sec %CP /sec %CP /sec %CP 50:10000:100/1000 21932 99 41475 100 10757 24 22114 99 45084 99 31565 99 ext3: Version 1.02a ------Sequential Create------ --------Random Create-------- testMX -Create-- --Read--- -Delete-- -Create-- --Read--- -Delete-- files:max:min /sec %CP /sec %CP /sec %CP /sec %CP /sec %CP /sec %CP 50:10000:100/1000 2168 21 45864 99 6555 28 3443 32 45943 100 144 1 reiser: Version 1.02a ------Sequential Create------ --------Random Create-------- testMX -Create-- --Read--- -Delete-- -Create-- --Read--- -Delete-- files:max:min /sec %CP /sec %CP /sec %CP /sec %CP /sec %CP /sec %CP 50:10000:100/1000 3208 87 10319 33 9505 94 3271 89 586 2 608 10 As you can see the random delete results on the ext2 drive are way better than reiser and ext3, and this is something a mailserver does on a regular basis when serving pop-boxes (especially if a user has say 2000 new mails in his mailbox and all he does is a "stat" and then disconnects the popd has to move all the mails from the "new" to the "cur" dir). Can you give me some insight why there are so big differences in the random delete part? best regards, michael
On Apr 05, 2002 01:29 +0200, Michael Renner wrote:> The test-partitions are on a ICP Vortex GDT8523RZ which runs in a RAID10 > setup, the filesystems got created/mounted with standard settings (except > 1024 byte blocksize on ext2/3 partitions). The noatime mount option hardly > affected the performance, switching from ordered to writeback in ext3 also > didn't change the results notable. I didn't dare to play around with > reiserfs mount options because i hardly have any knowledge about their > drawbacks.I would suggest benchmarking two things (independently and together): 1) running ext2/ext3 with a 4kB block size. 2) running ext3 with data=journaled While the bonnie++ benchmark may not show good results with #2, you will find that ext3 with data=journaled + sync I/O from MTAs has a very good performance. You will probably need to make the journal larger in this case (4x as large as the amount of email you will handle in 5 seconds), or 128MB-256MB or so. I would suggest using something like "postal" to benchmark a mail server.> As you can see the random delete results on the ext2 drive are way better > than reiser and ext3, and this is something a mailserver does on a regular > basis when serving pop-boxes (especially if a user has say 2000 new mails > in his mailbox and all he does is a "stat" and then disconnects the popd > has to move all the mails from the "new" to the "cur" dir).It's a journaling issue (ext3 actually writes to disk for deletes, while ext2 may skip writing a lot of stuff that has been invalidated before it needs to be written). Cheers, Andreas -- Andreas Dilger \ "If a man ate a pound of pasta and a pound of antipasto, \ would they cancel out, leaving him still hungry?" http://www-mddsp.enel.ucalgary.ca/People/adilger/ -- Dogbert