On Wed, 24 Oct 2007, Josh Carroll wrote:
Your tests with ffmpeg threads vs processes probably is triggering more
context switches due to lock contention in the kernel in the threads case.
This is also likely the problem with some super-smack tests. On each
context switch 4BSD has an opportunity to perfectly balance the CPUs. ULE
does not because it's too costly and hinders other workloads.
Thanks for the response Jeff, I appreciate it. Is this something that
the scheduler can recognize and auto-tune itself for? E.g. have it
recognize the scenario differences and do what 4BSD is doing for cases
such as ffmpeg and what ULE otherwise does in other circumstances? I
guess this would not come without its own overhead, which might defeat
the purpose anyway.
The problem is the two models are drastically incompatible. Rather than
use a shared queue for some threads I'm just going to have to continue to
refine the load balancing algorithm. The higher idle time with ULE is a
symptom of assinging work to the wrong place. I think I have a good idea
of how to improve that cpu selection.
I don't doubt that we can improve things further. It will just have to
wait for another few weeks before I'm able to do much about it.
Thanks again, I do appreciate your work and help. So you're
anticipating nothing can be done at all, or did you mean at the moment
until you get your equipment back and get settled? Or is this
something that would need to be looked at in -current rather than
7-STABLE?
I'm confident that we can improve things. It will probably not make the
cut for 7.0 since it will be too disruptive. I'm sure it can be
backported before 7.1 when ULE is likely to become the default.
I hope that we can continue to work together to verify any fixes I may
come up with.
Thanks,
Jeff
Regards,
Josh
_______________________________________________
[email protected] mailing list
http://lists.freebsd.org/mailman/listinfo/freebsd-performance
To unsubscribe, send any mail to "[EMAIL PROTECTED]"
_______________________________________________
[email protected] mailing list
http://lists.freebsd.org/mailman/listinfo/freebsd-performance
To unsubscribe, send any mail to "[EMAIL PROTECTED]"
