On 10 Apr 2007, at 12:40, John Madden wrote:
Indeed, I've seen reiserfs do some very long fsck's, but I've seen the same out of ext3. But for a filesystem of 35 million mail files, I figure it's got to beat ext3 on performance, at least. ...But there don't seem to be any stats at this scale to support that.
I've repeatedly seen RAID failures, etc, cause reiserfs to be meaningfully unrecoverable. I've never seen anything like that on ext3, but it's only been a couple of years. At the time we moved off reiserfs, ext3 also appeared to be better supported. We routinely reevaluate filesystem choice when we purchase new hardware, typically every two years. Since we're in the process of spec'ing new hardware, we'll be reviewing the filesystem choice soon.
As far as relative performance is concerned, as I recall properly tuned reiserfs performs slightly better than properly tuned ext3. Moving mailbox.db, seen DBs, etc, to their own IO chain does more for performance than anything else. I haven't experimented with running with three independent IO channel, but I gather that putting cyrus.cache and cyrus.index on their own partitions improves performance significantly as well.
Regarding scale, the backend I'm looking at just now has 9.8K user, 2.3TB of user data in 44M files. This is one of 23 backends in this murder. Generally speaking, I think replication is the way to go in either case.
:wes ---- Cyrus Home Page: http://cyrusimap.web.cmu.edu/ Cyrus Wiki/FAQ: http://cyrusimap.web.cmu.edu/twiki List Archives/Info: http://asg.web.cmu.edu/cyrus/mailing-list.html
