3 Oct
2007
3 Oct
'07
4:03 p.m.
On Wednesday 03 October 2007 15:05, Florian Schmidt wrote:
On Wednesday 03 October 2007, Wolfgang Woehl wrote:
I started this thread, but at the moment I'm not having much success at
getting realtime to work on Debian Etch, and that's before I start on Lenny.
Realtime works fine on Fedora 7, using the realtime low latency kernel from
planetccrma, plus rtirq, and a pam update from planetccrma that gave me a
limits.conf.rpmnew file, as below.
# limit realtime and memory locking access to users in the group audio
# there is no way to say "allow locking all memory", 4G should be enough
#
#* - rtprio 0
#* - nice 0
#
#@audio - rtprio 99
#@audio - nice -10
#@audio - memlock 4000000
# or (default) allow everyone access
* - rtprio 99
* - nice -10
* - memlock 4000000
Editing the existing limits.conf, adding the above, and commenting out the
default entries for jack, works fine. When starting jackd in qjackctl, I'm
getting one xrun every 6m 25secs on average, and the largest xrun over the
time I tested was about 0.325 msecs. I tested over about 2 1/2 hours, and the
only other apps open on the desktop were Xawtv, and Gkrellm.
On Debian Etch, after adding Marcos's repo to /atc/apt.sources.list, I
installed the 2.6.21-rt4, and 2.5.21.5-rt18 kernel, along with rtirq version
20070101. I edited limits.conf, and added the part above that I received from
planetccrma. That didn't work, with xruns every 2-3secs. I commented out the
default uncommented lines, and uncommented the audio lines, which are at the
top of this post. Still I get xruns every 2-3 secs, and if I reboot and try a
non rt kernel, xruns every 2-3 secs again. It appears that there is still
something missing to get realtime working on Etch.
I may have to download Musix again, and see how realtime runs on it. I DL'd
before, but that was when there was a problem with it booting with some
hardware. As I'm on dialup, I'd put off DL'ing it again, since the booting
problem has been resolved.
Marcos. If your reading this. Could you provide the link to download Musix
again please.
Thanks for all the replies to this thread.
Nigel.
Montag, 1. Oktober 2007 Marcos Guglielmetti:
Yes,
one always assumes with settings like this, that the programs which are run
are nicely behaved (e.g. jackd has a watchdog to kill its clients when they
run havok)..
But whether you give one process RT prio 1 or 99 doesn't matter. It can
always keep all SCHED_OTHER threads [the rest of your system, X, xterms,
whatnot] from running at all effectively freezing the box (though in
reality this is expected behaviout).
A typical method to counter this scenario would be to limit user processes
to prio 98 and have another watchdog process runnning at prio 99. That
watchdog would have another thread running at SCHED_OTHER and checks at
regular intervals whether this thread got to run at all. If not -> take
action and make e.g. all SCHED_FIFO threads in the system SCHED_OTHER so
the user is able to recover from the situation.. There are several
watchdogs available (including one from yours truly which needs an overhaul
though)..
And yes, of course, you wouldn't want to give all system memory away to
mlocking. You need some reserve..
Nice -10 otoh is pretty uncritical i guess.
Regards,
FLo @audio - rtprio 99
@audio - nice -10
@audio - memlock 4000000
Marcos, looking at these numbers (and variations of these on various
netpages) I always wonder whether they are such a good idea: Wouldn't
rtprio 99 allow for banzai processes to almost lock down the machine?
Wouldn't memlock 4000000 (4GB) be like begging for a swapfest on standard
1GB/2GB machines? Isn't the whole point about limits.conf to effectively
prevent processes from sucking in all resources? So that when realtime
processes go havoc you'll be left with cycles and memory to survive the
situation?