kernel preemption...
Warren Togami
warren at togami.com
Sun Apr 25 07:31:38 UTC 2004
Craig Cruden wrote:
> I noticed the latest kernels that have been
> distributed for testing have had the "PREEMPT" flag
> turned off. The pre-emption was added for things like
> sound -- i.e. desktop applications.... when the
> product is distributed is there going to be a
> "desktop" kernel version with it turned on by default?
I am nowhere near a kernel expert, so I do not know the details for
certain, but somebody explained to me that the upstream kernel still
does not fully have all race conditions fixed and preemption is still
known to cause crashes with certain drivers quite readily. So that is a
safety reason. Also some kernel devleopers have said that there is "NO
BENEFIT" (their emphasis) in enabling it, because scheduling latencies
are already much better in our latest kernel. If you study the
scheduler you will see that "fairness" for real-time processes is
dramatically improved by default. Anecdotally, xmms for the first time
no longer causes any skips in sound for me even during heavy disk and
processor usage.
Unfortunately some cases that I have seen the latest kernel fail with
scheduling latencies seem to be the fault of specific drivers for
specific hardware. Such problems are upstream issues and probably will
not be fixed unless somebody cares enough to dedicate time to it.
In any case, nothing stops you from building your own kernels. It may
even be preferred depending on your unique situation. With Open Source
you have the CHOICE, and be thankful for that.
>
> Also, a side question -- what are the benefits of the
> "4G/4G" changes (simple explaination) -- right now I
> am unable to use it turned on on my portable -- so I
> was wondering what I am missing out on :p
>
I was told that 4G/4G enables i686 to work much better with certain
types of software (I don't remember the details), and with the more
recent vdso patches added to the kernel much performance is regained.
What specifically is the issue you are seeing? Is it one of the two or
so known reported issues? Please provide very specific details and
testing information.
Warren
More information about the fedora-test-list
mailing list