Reproduceable Bug in reported GPU-Utilization

Graphics / Linux Linux
1

I found out how to reproduce and trigger the GPU-Utilization 0% bug.

System: Ubuntu 16.04
Driver: 367.27
GPU: GTX 1080

First, see the following nvidia-smi output.
My GPU is idle and the GPU-Utilization shows 0%.
So far so good.

root@ht:~# nvidia-smi
Sat Jun 18 14:52:08 2016
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 367.27                 Driver Version: 367.27                    |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  GeForce GTX 1080    Off  | 0000:01:00.0      On |                  N/A |
| 40%   53C    P0    35W / 180W |     10MiB /  8112MiB |      0%   E. Process |
+-------------------------------+----------------------+----------------------+

Now on some other terminal I’ll run some compute intense task.
This task causes the GPU to run at 100%.
It’s expected the GPU-Utiliazion to show 100%.
As you can see, it’s doing that, which is exactly how it should be.

root@ht:~# nvidia-smi
Sat Jun 18 14:57:31 2016
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 367.27                 Driver Version: 367.27                    |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  GeForce GTX 1080    Off  | 0000:01:00.0      On |                  N/A |
| 27%   62C    P2   189W / 217W |   1027MiB /  8112MiB |    100%   E. Process |
+-------------------------------+----------------------+----------------------+

Now how to trigger the bug. I’ll change the following line in my xorg.conf:

Original:

Option         "RegistryDwords" "PerfLevelSrc=0x2222"

New:

Option         "RegistryDwords" "PowerMizerEnable=0x1; PerfLevelSrc=0x2222"

So all I do here is to enable PowerMizer.
After the change in xorg.conf I have to restart both X11 and my compute intense.
The bug now occurs.
The GPU-Utilization will be shown as 0% even if it running at 100% in real:

root@ht:~# nvidia-smi           
Sat Jun 18 15:06:30 2016       
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 367.27                 Driver Version: 367.27                    |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  GeForce GTX 1080    Off  | 0000:01:00.0      On |                  N/A |
| 69%   62C    P2   196W / 217W |   1027MiB /  8112MiB |      0%   E. Process |
+-------------------------------+----------------------+----------------------+

As a proof that my compute intense task is running see reported power consumption.
Note that I was unable to trigger the bug by setting PowerMizer via nvidia-settings.
Using nvidia-settings I can enable PowerMizer without triggering the bug.
It needs to be done via xorg.conf.

–
atom

2

Hi atomu, I think its not recommended to use these RegistryDwords. As soon as issue hit please provide nvidia bug report by running nvidia-bug-report.sh script as super/root user. What compute intense application you are running? Can you share app or this issue can be reproduce with any other apps too?

3

Hey sandip, thanks for your reply. As requested I’ve recreated the case and when the bug hits in created the nvidia-bug-report.log.gz using nvidia-bug-report.sh and send to linux-bugs@nvidia.com.

The application I’ve used is my own software, it’s open-source software. You can clone the repository from GitHub: GitHub - hashcat/hashcat: World's fastest and most advanced password recovery utility and recreate the case locally. But I don’t think you need to use my software, it will make no difference which software you use.

4

Ticket number 200220143 for reference.