[ZODB-Dev] Use of fsync in FileStorage
Tim Peters
tim at zope.com
Thu Jul 29 23:32:15 EDT 2004
[Christian Robottom Reis]
> I made some extra timings with vmstat output for the run on the server I
> reported on before (raid-1 on lvd 10K scsi drives/adaptec 29160):
>
> No fsyncs: 8.29973 seconds, 1204.86 txn/sec
...
> One fsync: 11.4103 seconds, 876.397 txn/sec
...
> Two fsyncs: 12.8292 seconds, 779.474 txn/sec
That's "the other" Linux pattern emerging, which I'm associating with SCSI
drives (because I have to guess *something* <wink>):
- Less than a factor of 2 difference 0-vs-2 fsync (the other Linux pattern
is ballpark 10x difference here)
- Minor difference 1-vs-2 (in common w/ the other pattern)
- Most of the "damage" done by the first fsync (also in common)
> Doesn't really seem to make *that* much of a difference;
Right, I think even 0-vs-2 is minor on that box. You're not throwing 45K
txn/minute at your FileStorage in real life (do correct me if I'm wrong
about that, but I'd be majorly surprised if you were), and I bet the 2-fsync
rate is way more than enough for what you do throw at it. You're already
paying for 1 fsync in FileStorage, and the 1-vs-2 difference is somewhere
between trivial and minor here.
> this may have to do with VFS fsync behaviour when using raid-1, or
> perhaps the write behaviour on these specific hard drives --
Or all of the above, plus more. We're short on disk details in several of
the other reports.
> I am surprised so many others saw so much fluctuation in times.
Ah, you started using computers this year <wink>?
> I might run this on a similar server that I have remote access to if Tim
> says he wants me to.
If you're still willing, I'm still eager! Thanks.
More information about the ZODB-Dev
mailing list