If I understand it correctly, fine-grained file(descriptor) locking gave them a HUGE boost in
the MySQL/Postgres benchmarks. Last time I checked, file locking was still using the BKL under
Linux, so it might be good idea to attack this problem (any volunteers? ;)
Posted Feb 29, 2008 13:44 UTC (Fri) by andikleen (guest, #39006)
[Link]
That "last time" must have been at or before Linux 2.0 which came out in ~1996. Linux hasn't
used BKL in significant parts of the VFS (except perhaps
ioctl/fcntl) for a long time.
FreeBSD 7.0-RELEASE available
Posted Feb 29, 2008 18:34 UTC (Fri) by ballombe (subscriber, #9523)
[Link]
Is not fcntl precisely used for file locking (F_SETLK/F_GETLK) ?
FreeBSD 7.0-RELEASE available
Posted Feb 29, 2008 19:32 UTC (Fri) by cyrus (guest, #36858)
[Link]
Jonathan mentioned it in his 2004 article "The BKL lives on".
Quoting "The file locking code still requires the BKL."
http://lwn.net/Articles/86859/
FreeBSD 7.0-RELEASE available
Posted Feb 29, 2008 19:49 UTC (Fri) by cyrus (guest, #36858)
[Link]
I guess what i meant is the "global file list lock". Peter Zijlstra posted a patch in January
2007 but it wasn't accepted. See http://lkml.org/lkml/2007/1/28/29.
Along the lines Ingo Molnar posted some benchmarks: http://lkml.org/lkml/2007/1/28/116 You can
clearly see the improvements. Maybe Peter can start another attempt to get this upstream.