On 03/03/12 10:59, Adrian Chadd wrote:
Right. Is this written up in a PR somewhere explaining the problem in
as much depth has you just have?
Have no idea. I am new at this area and haven't looked on PRs yet.
And thanks for this, it's great to see some further explanation of the
current issues the scheduler faces.
By the way I've just reproduced the problem with compilation. On
dual-core system net/mpd5 compilation in one stream takes 17 seconds.
But with two low-priority non-interactive CPU-burning threads running it
takes 127 seconds. I'll try to analyze it more now. I have feeling that
there could be more factors causing priority violation than I've
described below.
On 2 March 2012 23:40, Alexander Motin<[email protected]> wrote:
On 03/03/12 05:24, Adrian Chadd wrote:
mav@, can you please take a look at George's traces and see if there's
anything obviously silly going on?
He's reporting that your ULE work hasn't improved his (very) degenerate
case.
As I can see, my patch has nothing to do with the problem. My patch improves
SMP load balancing, while in this case problem is different. In some cases,
when not all CPUs are busy, my patch could mask the problem by using more
CPUs, but not in this case when dnets consumes all available CPUs.
I still not feel very comfortable with ULE math, but as I understand, in
both illustrated cases there is a conflict between clearly CPU-bound dnets
threads, that consume all available CPU and never do voluntary context
switches, and more or less interactive other threads. If other threads
detected to be "interactive" in ULE terms, they should preempt dnets threads
and everything will be fine. But "batch" (in ULE terms) threads never
preempt each other, switching context only about 10 times per second, as
hardcoded in sched_slice variable. Kernel build by definition consumes too
much CPU time to be marked "interactive". exo-helper-1 thread in
interact.out could potentially be marked "interactive", but possibly once it
consumed some CPU to become "batch", it is difficult for it to get back, as
waiting in a runq is not counted as sleep and each time it is getting
running, it has some new work to do, so it remains "batch". May be if CPU
time accounting was more precise it would work better (by accounting those
short periods when threads really sleeps voluntary), but not with present
sampled logic with 1ms granularity. As result, while dnets threads each time
consume full 100ms time slices, other threads are starving, getting running
only 10 times per second to voluntary switch out in just a few milliseconds.
On 2 March 2012 16:14, George Mitchell<[email protected]> wrote:
On 03/02/12 18:06, Adrian Chadd wrote:
Hi George,
Have you thought about providing schedgraph traces with your
particular workload?
I'm sure that'll help out the scheduler hackers quite a bit.
THanks,
Adrian
I posted a couple back in December but I haven't created any more
recently:
http://www.m5p.com/~george/ktr-ule-problem.out
http://www.m5p.com/~george/ktr-ule-interact.out
To the best of my knowledge, no one ever examined them. -- George
--
Alexander Motin
_______________________________________________
[email protected] mailing list
http://lists.freebsd.org/mailman/listinfo/freebsd-hackers
To unsubscribe, send any mail to "[email protected]"
--
Alexander Motin
_______________________________________________
[email protected] mailing list
http://lists.freebsd.org/mailman/listinfo/freebsd-hackers
To unsubscribe, send any mail to "[email protected]"
