> > I'm curious about how your changes (merged and unmerged vmlocking  =

> > branch) impact performance on SMP machines.
> > I think that these changes are mostly positive :), but I'm not sure=
.=

>
> Merged: select/poll are slightly more efficient, system V semaphores
> do not require the kernel lock (although I need to switch that on).
> =

> Eventually: the VM system, ffs and tmpfs file systems, pipes, and a g=
ood
> chunk of other stuff will run without the kernel lock. Theoretically =
it will
> improve performance for a number of tasks on MP systems but the proof=
 will
> be in the numbers.

I did two full builds of yesterday's source tree (after build
fixes) freshly booted from the respective kernels, and this is
what I got:

host: NetBSD 4.99.22 (GENERIC.MP) #0: Wed Jul  4 12:01:27 CEST 2007
31896.071u 12360.153s 3:47:29.80 324.2% 0+0k 14202+2719081io 17567pf+13=
2w

host: NetBSD 4.99.23 (GENERIC.MP) #1: Tue Jul 10 18:44:20 CEST 2007
31708.518u 12436.039s 3:48:16.38 322.3% 0+0k 14285+2719957io 17566pf+13=
4w

I.e. the difference appears to be in the noise for now.  That may
be expected because the "Eventually" bits mentioned above aren't
in yet.

This was done on a 2x physical-CPU 2.2GHz Xeon with HT, so with 4
virtual CPUs, the build was done with -j 6 in both cases.

Regards,

- H=E5vard