Re: ULE Interactivity perf patch

From: Peter Schultz <pmes_at_bis.midco.net>
Date: Sat, 20 Dec 2003 08:43:33 -0600
Jeff Roberson wrote:
> On Fri, 19 Dec 2003, Peter Schultz wrote:
> 
> 
>>Jeff Roberson wrote:
>>
>>>I realized a pitfal in the way that I'm doing slice assignment for
>>>interactive tasks.  I'd like to have as many people test this as possible,
>>>in case there are unintended consequences.  What this patch does is allow
>>>interactive tasks to have longer time-slices so that they may be more
>>>efficient.
>>>
>>>This patch is intended to fix the poor performance of some interactive
>>>processes while under high load, especially high load with other
>>>interactive tasks present.
>>>
>>>http://www.chesapeake.net/~jroberson/interact.diff
>>>
>>
>>On this dual PII 350 box `make -j 11 buildworld', playing an mp3 with
>>xmms and unarchiving two separate mozilla distros can cause quite an
>>interactivity problem.  At a more moderate system load things are quite
>>usable.  It's still not BeOS, but BeOS never had the kind of i/o that
>>FreeBSD has, so I think things are going fairly well.
> 
> 
> Do these comments apply to ULE with and without the patch?  Is there any
> difference?  I suspect that your interactivity problems in this situation
> are more due to disk and memory pressure.  If you were to fire up vi, or a
> shell, something that's totally memory resident, do they suffer any lag?
> Also, how well does 4BSD do in this same test?
> 
Playing an mp3, I notice less and less of the skipping and popping
as I introduce your's and then Mathew Kanner's patches.

Running ULE with the latest Gnome, using the panel to select a window 
that is on a different workspace will pop the window up quite quickly, 
followed in due time by the rest of the desktop.  4BSD takes a few ms 
longer to complete this type of move because it wants to display the 
selected window along with the rest of the desktop it resides in, all in 
one step.

If I play two instances of xmms, run mozilla and type text in vi, i/o is
quite smooth with either scheduler.  If I add `make -j 2 buildworld' to 
the mix, there are widely spaced out glitches, which may occur a tish 
more often using 4BSD, but if I `tar yxf mozilla-source-1.0.2.tar.bz2' 
or `rm -rf mozilla', system i/o gets very out of tune regardless of the 
scheduler in use.  Typing in vi is interupted in sync with the audio 
bluurps.  I do find this situation slightly more tolerable with ULE.

With ULE there are less major skips, and it looks as though it's able to 
distribute tasks more evenly.  I'd say it's time to switch to ULE, but 
I'm just one guy testing on one machine.  I think progress is being made 
with ULE, because mp3 skips are more of a bluurp now, as apposed to the 
long hard scratches with 4BSD and older versions of ULE.  Playback slows 
rather than fails.  It seems to me that this is good.

Pete...
Received on Sat Dec 20 2003 - 12:50:48 UTC

This archive was generated by hypermail 2.4.0 : Wed May 19 2021 - 11:37:35 UTC