[ZODB-Dev] Use of fsync in FileStorage

[Tim Peters]

[Shane Hathaway]
> Here are my results for 10,000 transactions on a 1 GHz Athlon, Linux
> desktop using Reiserfs on a 40 GB drive.

IDE?

> One fsync: 76.3015 seconds, 131.059 txn/sec
> Two fsyncs: 86.915 seconds, 115.055 txn/sec
> No fsyncs: 9.38018 seconds, 1066.08 txn/sec

So this is one common <heh> Linux pattern so far:

- 10x difference 0-vs-2
- little difference 1-vs-2
- most "damage" done by the first fsync

> The first two tests made the hard drive work continuously.  In the third
> test, the hard drive sat idle except for a commit every 5 seconds.

At that level, impossible to distinguish from my Windows laptop.  In the
0-fsync test, I saw that Windows kept trickling data to disk even minutes
after the test ended (but I boosted the # of transactions to 100,000 for
that, else it all went too fast to get a feel for it by eyeball).

> If you're interested in the results, I could also try this test on a box
> I've been playing with: a dual 2.4 GHz Xeon hyperthreaded SuSE box with 2
> GB RAM and 4 SATA drives in a RAID-0 configuration.  It can write 180
> MB/s to its 900 GB array.  It screams like no other.  (Literally--the
> fans are really loud!)

I would love to see results on that.  That's deep in server-class territory,
and you may have noticed that major Zope customers seem to want the system
to run without problems <wink>.

> Or I could try it on the Gentoo box that actually runs Zope, a 333 MHz
> Pentium II with 128 MB RAM and 11 GB hard drive space. :-)

I'm truly interested in all the results we can get here.  I don't want you
to risk losing your family over it, but so long as you find it interesting
enough to run them, I'm eager to collect the results.