Re: ULE/yielding patch for testing.

On Sun, 7 Oct 2007 at 11:54:08 +0200, Marc Fonvieille wrote:
>On Sun, Oct 07, 2007 at 01:35:50AM -0700, Jeff Roberson wrote:
>>>> [...]
>>>> Not really when the slowness does not exist with 6.X.
>>> 
>>> I have experienced certain websites causing firefox to slow down 
>>> considerably.  When I mouse over firefox the mouse begins to jerk as well 
>>> but when I move back away from firefox it seems fine again.  I suspected 
>>> it was an x/firefox bug.
>> 
>> Actually I double-checked this case with dailytech.com.  It seems to be 
>> spending all of the cpu time in user-space in xwindows.  So it seems that 
>> it is the new xwindows that has the problem.  If you watch the 'TIME' field 
>> of top you can see that very little cpu time is attributed to firefox and 
>> quite a lot to X windows.  So the windowing system is too busy to redraw 
>> mouse events even if we are scheduling it well.
>> 
>
>You mean Xorg 7.2 or 7.3 with -CURRENT?
>Currently I observe the problem with Xorg 7.2 or 7.3 under -CURRENT,
>when Xorg 7.2 with 6.2-STABLE is Ok (I did not try with Xorg 7.3 on
>-STABLE, but I'm quite sure I'll end to same result).
>On a completly different machine (a more recent one with a 2.8GHz
>Celeron), the problem could be less obvious cause of the machine speed
>but I can notice the difference between 6.2-STABLE and 7-CURRENT with
>any Xorg version.
>
>> Jeff
>> 
>>> 
>>> Although I also sometimes observe mysql hanging for a period of a half 
>>> second or so while running sysbench.  I wonder if we don't have a 
>>> threading library bug?  I know I observed this prior to the thread_lock 
>>> work so hopefully I didn't break it.
>>>
>
>I think a lot of people missed that problem cause they run fast
>machines.  For me it's very difficult to track down the issue cause I'm
>ignorant in these OS areas and the various logs and tools don't give
>obvious information.

I have also been suffering from *atrociously* bad interactive response
during load (for example compiling, pkg_add, etc). Setting kern.hz=100
in /boot/loader.conf fixed this for me. With hz=1000 the machine was
practically unusable. But, with hz=100 and cpu-intensive loads running
I can easily and quickly switch between windows, watch a movie, etc
with almost no noticeable difference in responsiveness (from an idle
system). The difference in responsiveness between hz=1000 and 
hz=100; however, is *very* noticeable.

I'm running Xorg7.3 with ULE on 8-CURRENT on an UP AthlonXp 1600+
with 512 MB RAM. 

Cheers.
-- 
Mike Makonnen         | GPG-KEY: http://people.freebsd.org/~mtm/mtm.asc
mmakonnen _at_ gmail.com | AC7B 5672 2D11 F4D0 EBF8 5279 5359 2B82 7CD4 1F55
mtm _at_ FreeBSD.Org     | FreeBSD - http://www.freebsd.org