Re: fsync() madness

[Sami Liedes <sliedes@xxxxxxxxx>]

On Sun, Apr 20, 2008 at 12:45:54PM +0200, Oswald Buddenhagen wrote:
> hmm, it was on kde-devel, in fact.
> http://lists.kde.org/?l=kde-devel&m=119255926531628&w=2
> i think this is the second one i had in mind:
> http://lists.kde.org/?l=kde-core-devel&m=108971672603299&w=2

Well, I read those threads and everyone there seems to think it's no
major performance hit. That and I'd like to repeat that if it's good
enough for pretty much every other piece of software, it had better be
good enough for KDE. The current behavior is not ruining performance
only for KDE but for everybody else too using the same fs.

Just for testing, I replaced the retval = fdatasync() (or something)
call in KTempFile by retval = 0, and the speedup is massive. If you're
like me, you may have wondered what the heck KDE is doing during that
splash screen. Now I think I know. It's doing redundant fsync()s,
without which the splash screen is barely visible for maybe a fifth of
a second. Ditto for Konqueror and pretty much every KDE application,
the speedup is truly tremendous. No even need to speak about laptop
issues here, where spinning up the disk is a big problem for power
consumption. And really the only KDE application/class I've found so
far where I can think the fsync() is good for anything is KShred.

Well, if the KDE developers' attitude to fsync() is that the current
practice is good[*], I guess I will have to try to persuade the
maintainers of My Favorite Distro that this is a bug and the fsync()
should be disabled. I'd hope that not be the case. If it is your view
that data safety on power loss or crash trumps performance in every
sense, perhaps you should try to persuade users to mount their
filesystems noasync rather than forcing the thing down their throats.
("Options are bad" was said in one post in the threads mentioned
above...)

Sorry and thanks,

	Sami


[*] I'm not alleging it is, I definitely hope it's not, but I guess I
    have to wait and see.

Attachment: signature.asc
Description: Digital signature

>> Visit http://mail.kde.org/mailman/listinfo/kde-devel#unsub to unsubscribe <<