[linux-audio-user] Low latency with 2.6.16.16 vanilla

[Wolfgang Woehl]

Tuesday 23 May 2006 19:13, Lee Revell:
> On Wed, 2006-05-24 at 10:28 +0200, Wolfgang Woehl wrote:
> > Hi, I want to try what Lee mentioned a couple of times recently: Low
> > latency audio performance with 2.6 mainline -- no mingo-patch, no rt-lsm
> > (athlon xp 2600+, asus a7v8x-x, hdsp. On a 2.4.26 with lck patches this
> > system has good lowlat performance, solid jackd with -p 64 -n 2, so the
> > hardware should be ok. Ah well, looong dropouts on deep reiserfs walks
> > that never show up in jackd's messages but that's hopefully another
> > story).
> >
> > Got 2.6.16.16 from kernel.org. If I understood Lee right I could expect a
> > jackd with -p 64 -n 2 to work just fine but it doesn't. Loads of xruns.
> > What am I missing? What can I do to find out?
>
> -p 64 -n 2 is pushing the envelope of what the mainline kernel can do.
> 128 or 256 should be solid.  It will depend on the hardware and driver
> set.

Run as root I get no xruns with a jackd -R -d alsa -p 64 -n2 
(2.6.16.16-vanilla), with one of my average sessions (~8 tracks with plugins) 
the average jackd message is like

load = 17.4066 max usecs: 207.000, spare = 1244.000

(Only with a 70 tracks ardour session there are occasionals but the disk 
settings in ardour.rc seem to make a difference, they shouldn't I guess.) 
This is great. Vanilla for me from now on.

>
> Reiserfs is a poor choice of filesystem for low latency.
>
> Also make sure you are in realtime mode - depending on the distro, the
> realtime LSM may still be required.  I guess you are using the PAM
> method to enable non-root realtime?

I hadn't done my homework and wasn't aware of the privilege separation between 
root and users. So I had tried without any of the 3 possible mechanisms (pam, 
set_rlimits, rt-lsm?). And the way things are that doesn't work, ok. Thanks 
for the hint, Lee.

Damn, it doesn't make sense to me. If I can configure the kernel to be or not 
to be preemptible then what is the separation good for? Isn't overhead 
because of preemption the only price tag?

-- 
Wolfgang