[LAT] [LAU] Are RT-patches needed anymore? (Was Re: >= 2.6.27 RT ETA?)

[Robin Gareus]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

krgn wrote:
> On Sun, Feb 1, 2009 at 3:08 AM, <hollunder at gmx.at> wrote:
> 
>> On Sun, 01 Feb 2009 04:42:40 +0100
>> Robin Gareus <robin at gareus.org> wrote:
>>
>>> -----BEGIN PGP SIGNED MESSAGE-----
>>> Hash: SHA1
>>>
>>> Fernando Lopez-Lezcano wrote:
>>>> On Fri, 2009-01-30 at 03:23 +0100, torbenh at gmx.de wrote:
>>>>> On Thu, Jan 29, 2009 at 03:41:48PM -0800, Fernando Lopez-Lezcano
>>>>> wrote:
>>>>>> On Tue, 2009-01-27 at 14:06 +0100, Peder Hedlund wrote:
>>>>>>> Quoting Ken Restivo <ken at restivo.org>:
>>>>>>>
>>>>>>>> And here is the next installment in the saga of trying to get
>>>>>>>> Ingo RT going on my Asus EEE.
>>>>>>>>
>>>>>>>> I successfully built and ran the 2.6.26.8-rt12 with the
>>>>>>>> alsa_seq patch. It ran.
>>>>>>>>
>>>>>>>> The problem is that neither the Ethernet (atl1e) or wireless
>>>>>>>> (rt2860sta) work. So I pretty much had to reboot back out of
>>>>>>>> it immediately.
>>>>>>> I've been running the standard kernel from openSUSE 11.0 on my
>>>>>>> Athlon 2000+ and can get down to at least 5.3ms latency on an
>>>>>>> Audiophile 2496 using the limits.conf "trick".
>>>>>>>
>>>>>>> Do people really need lower latencies for music purposes or are
>>>>>>> we just thinking "well, I needed the RT patch three years ago; I
>>>>>>> ain't stopping now" ?
>>>>>> It depends on your usage (this question seems to come up every
>>>>>> couple of months lately). The current kernels are much better in
>>>>>> low latency applications than three years ago. They are usable if
>>>>>> you don't require "low" latencies (64 or 128 x 2). What you get
>>>>>> also strongly depends on the hardware mix you have.
>>>>>>
>>>>>> If you want to use 64 or 128 frame periods (or less) you probably
>>>>>> will need at rt patched kernel in most cases. Then again if an
>>>>>> occasional xrun is not a problem then you would be fine with the
>>>>>> stock kernel.
>>>>> i am running with -p64 -n3 on an intel-hda with 2.6.28
>>>>> of course internal cards have the greatest potential for
>>>>> lowlatencies. so this might be unfair, compared to pci.
>>>> Hmmm, I'm not sure, the load on pci itself by a soundcard should be
>>>> nothing really hard. What would the internal card use? Would not
>>>> that be pci or pci express anyway?
>>>>
>>> surely they are.
>>>
>>> $ lspci  | grep Audio
>>> 00:1b.0 Audio device: Intel Corporation 82801G (ICH7 Family) High
>>> Definition Audio Controller (rev 02)
>>>
>>> The RT patch does two things:
>>> It allows to prioritize interrupts and it [almost] guarantees
>>> real-time scheduling for a dedicated process or thread.
>>>
>>> While the soundcard is low bandwith on the PCI bus, IRQ prio may still
>>> be required to override HDD and [sometimes] graphics I/O; at least
>>> when playing or recording many tracks. NTL, you can get a perfect
>>> x-run free system without the RT patch; you can just not rely on it
>>> to be as x-run free as a RT patched kernel ;)
>>>
>>>>> and i havent really seen xruns which i could not relate to some
>>>>> programm which wasnt RT-safe, and i am compiling stuff most of the
>>>>> day... though perhaps i am not pushing the DSP load hard enough.
>>>>>
>>>>> i did not even turn preemptible RCU on.
>>>>> the latency measurement instrumentation is also in 2.6.28 btw.
>>>> Well, that's very good news then!
>>>>
>>>> I think the last time I tried to use a non-preempt was 2.6.27.x
>>>> (maybe, I would have to double check). Playing 24 channels in
>>>> ardour would result in xruns, not very often but they would happen,
>>>> this is with 128x2 on an RME hdsp card runing on a quad core intel
>>>> system. I should try again with the latest available.
>>> I just booted into a vanilla 2.6.28.2 #1 SMP PREEMPT
>>>
>>> right, there's no realtime patch, yet running jackd at 64 * 3 @48kSPS
>>> on a HDA - ardour2 with 12 tracks, a couple of LADSPA effects and
>>> jamin (lots of CPU!) - there's no xruns yet!
>>>
>>> I'll be back in the studio in two weeks from now to test it with USB
>>> and 1394 devices. With <=2.6.24 kernels those were always working more
>>> reliably that the HDA so I don't expect problems there.
>>>
>>> BTW. with 2.6.28 I needed to
>>>
>>>  `echo -1 > /proc/sys/kernel/sched_rt_runtime_us`
>>>
>>> or edit /etc/sysctl.conf and add
>>>   sys.kernel.sched_rt_runtime_us = -1
>>>
>>> before JACK was able to acquire real-time privileges.
>> May I ask what that does? This value is 950000 on my system.
>> I think jack has rt privileges here on 2.6.28.2 but I'm not too certain.
>> I don't even know how to check that reliably.
>>
> 
> jackd would not start with -R if realtime permission can not be granted. I
> think you would know if there was a problem ;)
> I have to say I don't know either what this option is you are setting,
> Robin. Is that the granularity of the timer (in u-sec's)? Setting to -1
> seems to suggest that the highest possible value is used, although I'm only
> wildly guessing. Would be great to know!

indeed, jackd did not start unless run as root or until I changed
/proc/sys/kernel/sched_rt_runtime_us

It's the us (or .001% of CPU usage?!) above which FIFO and RR will be
throttled to recover from a run-away rt process. - I'm not exactly sure
why it affects jackd when run as "normal" user; maybe because with CPU
load above the default 95% rt can not be guaranteed.

There's a bit [off topic] but enlightening info at
http://kerneltrap.org/mailarchive/linux-kernel/2008/8/16/2960524

BTW. in sysctl.conf it's  not "sys.kernel..." but
  kernel.sched_rt_runtime_us = -1

robin
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.9 (GNU/Linux)

iEYEARECAAYFAkmGZ1AACgkQeVUk8U+VK0JleACdGtXQ+jNwHFTkAIsHlN2Lo0+t
7CYAnRHWNCxkQ3CtuvmhLteUH+E6RXNJ
=q9dt
-----END PGP SIGNATURE-----