18 Jun
2011
18 Jun
'11
5:56 p.m.
On Mon, Jun 13, 2011 at 05:38:11PM -0700, Fernando Lopez-Lezcano wrote:
Hi all,
I just built a threaded-irqs enabled 2.6.39.1 kernel and used
cyclictest to try to get an initial feeling of how good it might be.
Cyclictest tests scheduling delays of the kernel it is running
on[*]. These tests were run on a Fedora 13 T510 Lenovo laptop, with
the ugly binary Nvidia driver (caveat emptor).
In the menu:
2.6.39.1 with threaded irqs and PREEMPT active
2.6.33.13-rt31
2.6.34.8-rt31 (a forward port of rt31 to 2.6.34)
plain Fedora kernel (2.6.34.8-68.fc13.i686.PAE)
Now, before someone points it out, this is just a synthetic test
that measures scheduling latency, not a real audio application.
still... it provides some numbers.
and what i get seems to differ from what you get pretty heavily.
---( torbenh@siel )---( ~/cvs/rt-tests )---
-> uname -a
Linux siel 2.6.39 #53 SMP PREEMPT Wed May 25 22:11:00 CEST 2011 x86_64
GNU/Linux
---( torbenh@siel )---( ~/cvs/rt-tests )---
-> ./cyclictest -t1 -p 80 -n -i 10000 -l 10000 -m
policy: fifo: loadavg: 2.34 1.27 0.54 5/267 10411
T: 0 ( 7904) P:80 I:10000 C: 10000 Min: 8 Act: 19 Avg: 17 Max:
97
---( torbenh@siel )---( ~/cvs/rt-tests )---
its just a single run. with a kernel compile in the background.
and thats WITHOUT threadirqs
See below for details... (look at the "Avg:"
and "Max:" latency
values, those are in microseconds)
Unless I have done something wrong, it looks like we have a long way
to go with regards to scheduling latency in anything other than
2.6.33 + rt patches. But, with threaded irqs in 2.6.39 maybe the
performance with audio applications running under jack will be good
anyway as we will be able to tune irq processing. Seems to me
unlikely. I have not tested that yet.
you did not use -m ... although that didnt really change much for me.
but i have plenty of RAM here.
Tests were done after a clean reboot, initially with the machine
idle and then while loading thunderbird with tons of emails and
firefox with tons of pages. No audio work.
hmm... maybe i should have loaded up ff...
but i loaded midori. because ff was already running.
YMMV...
-- Fernando
[*] https://rt.wiki.kernel.org/index.php/RT_PREEMPT_HOWTO
----------------------------------------------------------------------
LENOVO T510
--
torben Hohn