NVIDIA dual monitors kde slow

[codelion]

Re: NVIDIA dual monitors kde slow

No go for Nvidia Quadro FX 3700M before 180, i.e. no 173.
See http://www.nvidia.com/object/linux_d...32_180.44.html

[codelion]

Re: NVIDIA dual monitors kde slow

Thought just using one screen works, and still got slow during work day. While it could happen various circumstances, a good way to make it happen was in a VMware VM to play CNN video feed full screen, which I don't do during work.

Well, maybe I have a workaround this time. Didn't get a slowdown during my test.

Below is the xorg.conf which seems to be working now.

I share it so you can try, but you have to do your own thinking whether this works in Kubuntu. I'm here because we have same hardware Dell Precision M6400 with Nvidia 2700M or 3700M, Linux, VMware and slowdown problem.

Section "Files"
	ModulePath "/usr/lib64/xorg/modules/extensions/nvidia"
	ModulePath "/usr/lib64/xorg/modules"
EndSection

Section "ServerFlags"
#	Option   "AIGLX" "on"
	Option   "AIGLX" "off"
EndSection

Section "Device"
	Identifier "Videocard0"
	Driver   "nvidia"
#	Option   "TwinView" "on"
#	Option   "TwinViewOrientation" "CRT-0 Above DFP-0"
	Option   "Coolbits" "1"
	Option   "TripleBuffer" "on"
	Option   "UseEvents" "true"
#	Option   "OnDemandVBlankInterrupts" "true"
#	Option   "VertRefresh" "DFP-0: 59.9; CRT-0: 60"
#	Option   "UseEdidFreqs" "false"
#	Option   "ModeValidation" "NoVertRefreshCheck"
	Option   "Accel" "false"
	Option   "NoMTRR" "true"
EndSection

Section "Extensions"
#	Option   "Composite" "Enable"
	Option   "Composite" "Disable"
EndSection

Oh, I'm running it at low GPU speed too.

GNOME → System → Preferences → Startup Applications
Startup Programs tab
Add button
Name: Nvidia Downclock
Command:

nvidia-settings --assign='[gpu:0]/GPUPowerMizerMode=1' --assign='GPUOverclockingState=1' --assign='GPU2DClockFreqs=200,100' --assign='GPU3DClockFreqs=300,200'

Comment: Set lower clock speed for Nvidia GPU

[codelion]

Re: NVIDIA dual monitors kde slow

Those settings didn't fix the problem. Slowness has struck again. Hence I ran

sudo powertop -d -t 10

and among other output got

Wakeups-from-idle per second : 688.2	interval: 10.0s
no ACPI power usage estimate available
Top causes for wakeups:
 60.3% (3172.9)   <kernel core> : hrtimer_start_range_ns (tick_sched_timer) 
 15.0% (789.3)    <interrupt> : extra timer interrupt 
 13.1% (687.1)    vmware-vmx : __mod_timer (LinuxDriverPollTimeout) 
 10.1% (531.0)    vmware-vmx : hrtimer_start_range_ns (hrtimer_wakeup) 
  0.3% ( 16.9)   <kernel IPI> : Rescheduling interrupts 
  0.3% ( 15.7)    <interrupt> : ahci 
  0.2% ( 11.2)      vmware : hrtimer_start_range_ns (hrtimer_wakeup) 
  0.2% ( 11.0)  vmware-unity-he : hrtimer_start_range_ns (hrtimer_wakeup) 
  0.1% ( 3.9)   <kernel core> : __mod_timer (rh_timer_func) 
  0.0% ( 2.6)   <kernel IPI> : TLB shootdowns

What exactly is the number in parentheses in "39.4% (115.6) foo: bar"? It's the number of wakeups per second and per CPU core. Per http://www.lesswatts.org/projects/powertop/faq.php.

Apparently that's really bad, 3000 of those wakeups <kernel core> : hrtimer_start_range_ns (tick_sched_timer) per second per CPU core. See https://bugs.launchpad.net/ubuntu/+s...ux/+bug/373245. Seems to affect Linux notebooks. Apparently a Dell Precision M6400 can produce large numbers.

Do we have a solution? Not yet, but good to see over there with powertop that others have similar problems. So now I'm searching along the line of what causes that interrupt.

Here is example output of

sudo powertop -d -t 10

when the problem has gone away (a bad VM closed):

Wakeups-from-idle per second : 262.3	interval: 10.0s
no ACPI power usage estimate available
Top causes for wakeups:
 42.8% (193.1)    <interrupt> : extra timer interrupt 
 32.1% (144.7)   <kernel core> : hrtimer_start_range_ns (tick_sched_timer) 
 10.5% ( 47.2)      firefox : hrtimer_start_range_ns (hrtimer_wakeup) 
  2.7% ( 12.3)    <interrupt> : ahci 
  2.6% ( 11.5)      vmware : hrtimer_start_range_ns (hrtimer_wakeup) 
  2.4% ( 11.0)  vmware-unity-he : hrtimer_start_range_ns (hrtimer_wakeup) 
  1.7% ( 7.8)   <kernel core> : hrtimer_start (tick_sched_timer)

Suggested reading man powertop and I'm now starting to read http://www.linuxpowertop.org/ and its http://www.lesswatts.org/projects/powertop/faq.php.

[codelion]

Re: NVIDIA dual monitors kde slow

I'm now also using

sudo more /proc/timer_stats

see http://www.mjmwired.net/kernel/Docum...imer_stats.txt and

sudo perf top

and from what I see and reading https://bugs.launchpad.net/ubuntu/+s...ux/+bug/145377 next thing I'm going to try running with kernel option noapic. That's correct, no Advanced Programmable Interrupt Controller (APIC). That isn't a typo of anything similar. I should report once I know.

[codelion]

Re: NVIDIA dual monitors kde slow

Quick note in the middle of a workday, solution seems to be working:

I've been running well for a while now, with several VMs in VMware 7.0.1 on this Dell Precision M6400 with quad core CPU with nVidia 3700M running a x86_64 Linux with a 2.6.32 kernel, by having in addition to above reply #16 mentioned xorg.conf (which presumably reduces GPU use) the following kernel options:

nomodeset (required for proprietary nvidia driver, open ones fail on stuff like resume from suspend)
iommu=soft (required for proprietary nvidia driver, something they don't do right)
noapic (required for stability, make sure you spell that exactly so, not something else)

The issue seems to occur on more than one flavor Linux (Ubuntu and Fedora), probably same kernel versions, and seems to be more of a problem on notebooks, possibly made worse by multiple cores (which most have) or that graphics card.

Also, the problem also occurred at least once with the nouveau driver, so you can't just blame the closed source driver.

[codelion]

Re: NVIDIA dual monitors kde slow

Not really fixed. Since using noapic kernel option I started getting a new kind of crash from VMware, at least twice, hence no I'm longer using noapic kernel option. Living with more than daily slowdown. Will try newer kernel 2.6.32.14 instead of 2.6.32.12, but only after I have time for total backup.

[dibl]

Re: NVIDIA dual monitors kde slow

I don't claim expertise with dual-monitor maneuvers, but here is a guy with your GPU, who says the new 256.35 driver is working well.

[codelion]

Re: NVIDIA dual monitors kde slow

Here is an interesting observation I seem to have made more than once:

If I have a VMware VM having that slows down, and quit it, and after so many minutes (10, 30, whatever) it finally it is shut down, and then I start the same VMware VM again, it still is slow. When, however, I restart the whole computer then it works again normally.

Could kgtboca confirm whether he experiences the same?

[dibl]

Re: NVIDIA dual monitors kde slow

a VMware VM having that slows down, and quit it, and after so many minutes (10, 30, whatever) it finally it is shut down

I have seen that behavior once -- it was due to a dying XFS filesystem. I hope you don't have anything like that going on. You might consider the security of running fsck on the partition where the VM lives.

[codelion]

Re: NVIDIA dual monitors kde slow

Had to vacuum the air flow openings, now it seems fixed.

After many efforts one day I noticed in powertop the Dell Precision M6400 with quad core CPU and Nvidia Quadro FX 3700M was stuck in lower speed of SpeedStep 1.6GHz. Found forum entry about vacuum cleaning.

Finally installed a GUI for temperature sensors. My previously observed idle temperature of 70C really was persistent. Sucked out dust from bottom of machine, surprisingly a cloud was blown out at back, and within seconds it was down to 60C and SpeedStep was back up at 2.5GHz and within a few more seconds CPU use percentage was back to reasonable. Within minutes down to 55C, 53C. Machine seems to perform much better now. I had visually inspected for dust in the slots and hadn't seen any, so I was surprised, but the GUI for temperature sensors made a point. Now at 51C, 52C.

Apparently this model hardware (Dell Precision M6400, maybe also Dell Precision M6500) is pushing the limits of heat to the point of impacting performance.

Bad airflow apparently has strange side effects, such as oddly amplified slowdown. Must be something like GPU switches into a funny mode when it gets hot.

Vacuum cleaning the air flow openings apparently has helped. Yay!

In hindsight, the pattern of failure and recovery was consistent with when one would expect overheating.

Also, I did remove the anti-static mat under the machine, it made the fan work too hard, it was a tad too soft.

While chasing this I did learn even more about optimizing VMs, also how to underclock Nvidia.

Also noticed Fedora 13 with kernel 2.6.33 uses way less interrupts than Fedora 12 did with kernel 2.6.32, seen per powertop. Hope newer Ubuntu has same improvements.