On Sun, Oct 26, 2003 at 05:38:29AM -0500, rwhron@earthlink.net wrote:
> 2.6.0-test8-mm1 doesn't have the AIM7 database regression
> that 2.6.0-test8 has.  The AIM7 fileserver and shared 
> workloads regression between test5-mm2 and test8-mm1 is
> about 18%.  Mainline regression was about 38% between test5
> and test8 on AIM7 fileserver and shared workloads.
> 

Yes, I've noticed this as well.  This has been the pattern since
test5.  Every mainline release (test6, test7, test8, AND test9)
has a large regression in as-iosched vs deadline scheduler performance.
The regression is considerably smaller when the mm[1-4] patches are
applied.  See the attachment for a collection of data points, collected
running REAIM, an updated version of AIM7.

In the mm[1-4] broken-out patches lists, there is an O_SYNC-SPEEDUP-2.patch
that has been in all of the mm[1-4] patches since test5.  The comments at
the head of that patch says this patch reduces the blocking of writers on i_sem.

This patch probably allows more writer concurrency.  This benefits
only O_SYNC writes.   Reaim does O_SYNC writes.  My guess this patch
is what reduces the as-iosched regression for mm[1-4] patched kernels.

I imagine the O_SYNC-SPEEDUP is unlikely to get into any more TEST kernels.
But maybe it gives a hint is to what's going on with as-iosched.
Letting in more concurrent writers seems to reduce as-iosched's peformance
regression.