FreeBSD 5.3 I/O Performance / Linux 2.6.10 | Continued Discussion

[Robert Watson]

On Thu, 27 Jan 2005, Robert Watson wrote:

> While it's not for the feint of heart, it might be interesting to see
> how results compare in 6-CURRENT + debugging of various sorts (including
> malloc) turned off, and debug.mpsafevfs turned on.  One possible issue
> with the twe/twa drivers is that they are currently MPSAFE, so may see
> substantial contention (and hence additional latency).  The move to an
> MPSAFE VFS will help with that a lot, I should think. 

And, if you're in the mood for hacking code, and promise not to use
snapshots, try making vfs_subr.c:vn_start_write(),
vfs_subr.c:vn_write_suspend_wait(), vfs_subr.c:vn_finished_write(),
vfs_subr.c:vfs_write_suspend(), and vfs_subr.c:vfs_write_resume() into
noop's.  These calls are used to avoid some deadlock scenarios associated
with snapshot generation, but they also introduce a small but non-trivial
amount of overhead to a number of operations.  Since you're set up to do
some testing, knowing how much of that cost is from these operations
should be quite interesting. 

Robert N M Watson