[mythtv-users] Heavy load troubles

[David Rees]

Brian Foddy wrote:

> Since Thanksgiving, I built up a new machine, a single 3Ghz xeon,
> 1GB ram, the same scsi matched 10K disks (now 4) in raid0, Nvidia FX5600,
> and the same 3 tuner cards.  By all reasonable comparisons, the new
> machine is many times the machine as the old.  I intentionally
> overbuilt to handle a pcHDTV card, which is currently installed
> in a test machine until this get more stable.

Out of curiosity, what motherboard are you using, and what NVidia drivers?

> Over the past month, I have seen almost all the following symptoms
> starting with Myth .12+ cvs, .13 release, back to .12 release,
> and now .13+ cvs from Dec 17.  I've upgraded several ivtv versions
> and am now using the 1125a release recommended for .13.  The machine
> is running Mandrake 9.2 (2.4.22mdk_smp) with current patches, and is 
> dedicated to myth so there are no other jobs running.

I've had issues with the 1125a release hanging very similarly to what 
you've described.  I've only got one PVR-250 in my machine.  Try the 
CVS-LATEST tarball on the ivtv website that has been working very well 
for me the last few days.

> 4.  Looking at "top" and xosview while running the 3 shows and playing
> one of them, I notice a very strange cpu load behavior, its not
> consistant, not even close.  Even using my old low-quality settings,
> there is an aproximately 4 minute cycle where the cpu load slowly goes from
> almost nothing up to 80-100%, followed by a sharp drop back to 5-10%
> again.  About 4 minutes later the cycle repeats.  Changing recording quality
> (even dramatically, 640x480 mpeg at 4500) doesn't change the behavior, 
> only the cycle period and the cpu starting level.  Top shows the backend
> and X server taking most of the cpu during the peak.

Do you have automatic transcoding turned on?  What is using up the CPU?

> 3.  I normally run the backend as root, to activate the renice
> logic of the card reading thread to -10.  
> I've even tried changing the source to make that
> reniced thread a real-time process (SCHED_RR).  No apparent help.

Shouldn't be necessary.

> 4.  Turned on and off (mostly off) the cpu Hyperthreading.  I'm always
> running a SMP kernel.

I would leave it off, and run a non-SMP kernel, Hyperthreading won't 
help performance on a MythTV type workload, and that's just another 
variable thrown into the mix.

-Dave