Re: ULE/SCHED_SMP diff for 7.0

On Wed, 18 Jul 2007, Dmitry Morozovsky wrote:

> On Mon, 16 Jul 2007, Jeff Roberson wrote:
>
> JR> http://people.freebsd.org/~jeff/ule.diff
> JR>
> JR> This patch is scheduled for inclusion in 7.0.  I would like anyone who cares
> JR> to run it to validate that it does not create any stability or performance
> JR> regression over the existing ULE.  This patch replaces ULE with SCHED_SMP,
> JR> which will now no longer exist as a seperate fork of ULE.
> JR>
> JR> Briefly, this is still a very suitable scheduler for uniprocessor machines
> JR> while providing stronger affinity and other performance improvements for
> JR> multiprocessor machines.
> JR>
> JR> Even "works for me!" type responses are welcome so I know roughly how many
> JR> people have tested before I commit this close to release.
>
> Works for me with MSI S420 lapton (core2duo T2400) on fresh -CURRENT/i386
> CPU: Genuine Intel(R) CPU           T2400  _at_ 1.83GHz (1833.41-MHz 686-class CPU)
>
> However, make -j3 buildworld buildkernel (trimmed GENERIC) time degraded by 7%:
>
> marck_at_mck-s420:/var/tmp> ministat before after
> x before
> + after
> +------------------------------------------------------------------------------+
> |   x                                                                    +     |
> |   x                                                                    +     |
> |x  x        x                                                      +    +    +|
> ||__MA____|                                                           |__A__|  |
> +------------------------------------------------------------------------------+
>    N           Min           Max        Median           Avg        Stddev
> x   5         36.33         36.77         36.45        36.486    0.16637307
> +   5         38.87         39.23         39.03        39.042    0.12774976
> Difference at 95.0% confidence
>        2.556 +/- 0.216322
>        7.00543% +/- 0.59289%
>        (Student's t, pooled s = 0.148324)
>
> before = kernel with SCHED_4BSD, after = SCHED_ULE with your patchset. Both
> are without WITNESS/INVARIANTS.

Hi Dmitry,

Can you test with http://people.freebsd.org/~jeff/sched_ule.c?  I believe 
this will somewhat improve the situation with buildkernel.  I have tested 
with my own core2duo laptop.  The first run after reboot is now about 7% 
faster than before.  Subsequent runs are not improved as much.  Only 2-3%. 
You can also try tuning sysctl kern.sched.steal_thresh down as low as 1 to 
see how much this may help.  Unfortunately lower values tend to really 
hurt other tests.

Thanks,
Jeff

>
> Sincerely,
> D.Marck                                     [DM5020, MCK-RIPE, DM3-RIPN]
> ------------------------------------------------------------------------
> *** Dmitry Morozovsky --- D.Marck --- Wild Woozle --- marck_at_rinet.ru ***
> ------------------------------------------------------------------------
> _______________________________________________
> freebsd-current_at_freebsd.org mailing list
> http://lists.freebsd.org/mailman/listinfo/freebsd-current
> To unsubscribe, send any mail to "freebsd-current-unsubscribe_at_freebsd.org"
>