On 10/9/2011 9:32 AM, Wietse Venema wrote: > Stan Hoeppner: >> On 10/8/2011 3:33 PM, Wietse Venema wrote: >>> That's a lot of text. How about some hard numbers? >> >> Maybe not the perfect example, but here's one such high concurrency >> synthetic mail server workload comparison showing XFS with a substantial >> lead over everything but JFS, in which case the lead is much smaller: >> >> http://btrfs.boxacle.net/repository/raid/history/History_Mail_server_simulation._num_threads=128.html > > I see no write operations, no unlink operations, and no rename > operations.
Apologies. I should have provided more links. The site isn't setup for easy navigation... >From the webroot of the site: http://btrfs.boxacle.net/ Mail Server (raid, single-disk) Start with one million files spread across one thousand directories. File sizes range from 1 kB to 1 MB Each thread creates a new file, reads an entire existing file, or deletes a file. 57% (4/7) reads 29% (2/7) creates 14% (1/7) deletes All reads and writes are done in 4 kB blocks. > Comments on performance are welcome, but I prefer that they are > based on first-hand experience, and preferably on configurations > that are likely to be seen in the wild. I would love to publish first hand experience. Unfortunately to sufficiently demonstrate the gains I would need quite a few more spindles than I currently have available. With my current hardware the gains w/XFS are in the statistical noise range as I can't sustain enough parallelism at the spindles. That said, there are plenty of mailbox servers in the wild that would benefit from the XFS + linear concat setup. It doesn't require an insane drive count, such as the 136 in the test system above, to demonstrate the gains, especially against EXT3/4 with RAID5/6 on the same set of disks. I think somewhere between 16-32 should do it, which is probably somewhat typical of mailbox storage servers at many sites. Again, this setup is geared to parallel IMAP/POP and Postfix local delivery type performance, not spool performance. The discussion in this thread drifted at one point away from strictly the spool. I've been addressing the other part. Again, XFS isn't optimal for a typical Postfix spool. I never made that case. For XFS to yield an increase in spool performance would likely require and unrealistically high inbound mail flow rate and a high spindle count to sink the messages. I'll work on getting access to suitable hardware so I can publish some thorough first hand head-to-head numbers, hopefully with a test harness that will use Postfix/SMTP and Dovecot/IMAP instead of a purely synthetic benchmark. -- Stan
