Hello everyone,
While I am not sure whether this is a bug or not, the phenomenon is
rather peculiar so I have decided to ask here. The problem is, upgrading
the kernel on any of my two AMD64 boxes from 2.6.24 to 2.6.25 or newer
causes both systems to get "hiccups", that it every half a second or so
they briefly freeze. Below you will find two examples of the problem:
1. On both systems, refreshing icons in the system tray takes much more
time than before: whereas with 2.6.24 the flicker of icons as they get
updated is barely noticeable, with 2.6.25+ you can see it quite clearly
and sometimes you can even see blank background of the tray app.
2. On the weaker of the two systems (Athlon64 3500+, 2 GB RAM,
nForce-based motherboard, on-board sound device supported by the
intel8x0 ALSA driver), one can hear audio playback (tested primarily
on a Shoutcast stream) stutter on a regular basis even with
buffering enabled. With 2.6.24 this never happened, even under heavy
system load.
3. I run an instance of Internet DJ Console (idjc), which used JACK, on
the stronger of the two systems (Athlon64 4000+, 2 GB RAM, nForce-based
motherboard, Sound Blaster Audigy 2 PCI card with ALSA driver emu10k1).
Up to and including 2.6.24 I've never had any problems with streaming
to the Icecast server I use, even under relatively heavy system load -
those few xruns, of just a few ms each, that did happen were removed by
buffering. However, under 2.6.25 and up I start getting xruns as soon
as idjc starts, even before I connect to the Icecast server or start
broadcasting anything! These look as follows:
**** alsa_pcm: xrun of at least 1222893019070.464 msecs
(the value is always the same) and appear more-or-less twice per
second. When I do connect to the server and begin streaming, the sound
as received by clients is audibly choppy when the system is otherwise
idle and becomes extremely distorted under load. The problem persists
after switching to an Icecast server running on the same machine, or indeed
if instead of running an independent media client I enable (in idjc)
listening to the stream before it gets sent to the server.
As I mentioned earlier, the problem affects all kernel versions I have
tried from 2.6.25 up to and including 2.6.26.5 - both vanilla versions
and with the grsecurity patch applied. Attached you will find my config
files for 2.6.24.7 and 2.6.26.5. Unfortunately I do not know whether the
problem affects other architectures than amd64.
I would greatly appreciate it if you could look into the problem and
let me know what is wrong - i.e. whether this is a bug, an
unnecessarily-enabled feature, something that just needs a bit of
tuning and so on. If you need any more information regarding the issue,
my set-ups and so on, please don't hesitate to contact me (by
writing to my e-mail address - I'm not subscribed to LKML at present)!
Yours sincerely,
--
MS
On Wed, Oct 1, 2008 at 2:51 PM, Marek Szuba <[email protected]> wrote:
> Hello everyone,
>
> While I am not sure whether this is a bug or not, the phenomenon is
> rather peculiar so I have decided to ask here. The problem is, upgrading
> the kernel on any of my two AMD64 boxes from 2.6.24 to 2.6.25 or newer
> causes both systems to get "hiccups", that it every half a second or so
> they briefly freeze. Below you will find two examples of the problem:
>
> 1. On both systems, refreshing icons in the system tray takes much more
> time than before: whereas with 2.6.24 the flicker of icons as they get
> updated is barely noticeable, with 2.6.25+ you can see it quite clearly
> and sometimes you can even see blank background of the tray app.
>
> 2. On the weaker of the two systems (Athlon64 3500+, 2 GB RAM,
> nForce-based motherboard, on-board sound device supported by the
> intel8x0 ALSA driver), one can hear audio playback (tested primarily
> on a Shoutcast stream) stutter on a regular basis even with
> buffering enabled. With 2.6.24 this never happened, even under heavy
> system load.
>
> 3. I run an instance of Internet DJ Console (idjc), which used JACK, on
> the stronger of the two systems (Athlon64 4000+, 2 GB RAM, nForce-based
> motherboard, Sound Blaster Audigy 2 PCI card with ALSA driver emu10k1).
> Up to and including 2.6.24 I've never had any problems with streaming
> to the Icecast server I use, even under relatively heavy system load -
> those few xruns, of just a few ms each, that did happen were removed by
> buffering. However, under 2.6.25 and up I start getting xruns as soon
> as idjc starts, even before I connect to the Icecast server or start
> broadcasting anything! These look as follows:
>
> **** alsa_pcm: xrun of at least 1222893019070.464 msecs
>
> (the value is always the same) and appear more-or-less twice per
> second. When I do connect to the server and begin streaming, the sound
> as received by clients is audibly choppy when the system is otherwise
> idle and becomes extremely distorted under load. The problem persists
> after switching to an Icecast server running on the same machine, or indeed
> if instead of running an independent media client I enable (in idjc)
> listening to the stream before it gets sent to the server.
>
>
> As I mentioned earlier, the problem affects all kernel versions I have
> tried from 2.6.25 up to and including 2.6.26.5 - both vanilla versions
> and with the grsecurity patch applied. Attached you will find my config
> files for 2.6.24.7 and 2.6.26.5. Unfortunately I do not know whether the
> problem affects other architectures than amd64.
>
> I would greatly appreciate it if you could look into the problem and
> let me know what is wrong - i.e. whether this is a bug, an
> unnecessarily-enabled feature, something that just needs a bit of
> tuning and so on. If you need any more information regarding the issue,
> my set-ups and so on, please don't hesitate to contact me (by
> writing to my e-mail address - I'm not subscribed to LKML at present)!
>
> Yours sincerely,
> --
> MS
>
$ bzgrep -i group config.2.6.26.5.bz2
# CONFIG_CGROUPS is not set
CONFIG_GROUP_SCHED=y
CONFIG_FAIR_GROUP_SCHED=y
# CONFIG_RT_GROUP_SCHED is not set
# CONFIG_CGROUP_SCHED is not set
The group scheduler has (or at least had, prior to now) known fairness
and interactivity issues. Please disable it and try again.
Marek Szuba <[email protected]> writes:
>
> I would greatly appreciate it if you could look into the problem and
> let me know what is wrong - i.e. whether this is a bug, an
> unnecessarily-enabled feature, something that just needs a bit of
> tuning and so on. If you need any more information regarding the issue,
> my set-ups and so on, please don't hesitate to contact me (by
> writing to my e-mail address - I'm not subscribed to LKML at present)!
You could do a oprofile run (http://oprofile.sourceforge.net)
and see if it's something taking a lot of CPU time.
-Andi
--
[email protected]
On Wed, 1 Oct 2008 15:09:40 -0700
"Ray Lee" <[email protected]> wrote:
> The group scheduler has (or at least had, prior to now) known fairness
> and interactivity issues. Please disable it and try again.
Your assumption was correct, disabling the group scheduler has fixed
all the problems (except for the fact that when xruns do occur the
values are still extremely large - but that's probably just version
incompatibility between ALSA drivers in the kernel and libasound). Of
course this suggests that the group scheduler is still borked, at least
on amd64.
Thanks for your help!
Best regards,
--
MS