[ardour-users] [Fwd: Re: Is 64 bit dual core supported and stable yet?]

[Fernando Lopez-Lezcano]

On Sat, 2005-12-03 at 23:33 -0500, Joel White wrote:
> Fernando Lopez-Lezcano wrote:
> 
> >On Sat, 2005-12-03 at 10:49 -0500, Joel White wrote:
> >>> 
> >>><Snip>
> >>> 
> >>I am currently running stock 2.6.14 with the 3800+ X2 (that's me, 
> >>below).  If I set the smp_affinity of my sound card to one processor, 
> >>the "delay of xx usecs exceeds.." messages virtually disappear.  Without 
> >>that, they pop up regularly, with increasing delay times.
> >> 
> >Yup, that's the problem of drifting TSC's. Hmmm, most probably you want
> >to pin the jack processes to one processor, not the soundcard. 
> >  
> Ok, cool.  I was wondering about that.  I've just tried setting the cpu 
> affinity of the jack processes to one processor using taskset, along 
> with setting the smp_affinity of the sound card IRQ back to 0x03 (both 
> processors).  I've been recording 8 tracks at -r 44100 -p 64 -n 2 
> (duplex) for over an hour now without an xrun or "delay..." message.  I 
> think this will give me a good working system for recording.

Sounds very good. Did you reboot? The TSC drift depends on the cpu being
idle and for how long. So right after a reboot you won't see the problem
but it will get worse over time (without reboots and with light
workload). A loaded system will definitely drift very little. I have
note tested vanilla 2.6.14 but on -rt patched kernels the situation
could take a while to show up, once it did it kept getting worse over
time, of course. 

> >>As a test, I just got through recording 55+ min of 8 tracks (-r 44100 -p 
> >>64 -n 2, duplex) before ardour halted with an xrun and "delay.." error.  
> >>   
> >I sounds to me that -p64 is too low for the stock kernel in any case. 
> >
> I don't think so - the xruns and "delay..." values were way outside what 
> one would expect given the jack settings (i.e., tens to hundreds of msec).

Could be that vanilla 2.6.14 is better than what I thought. 

FYI the latest -rt patch (-rt21) does not choose the TSC for system
timing - that plus a patched jack should also be good without pinning
processes to individual cpus (not to suggest you switch to that). 

-- Fernando