I am trying to collect performance metrics of a GLSL compute shader on Linux (archlinux, kernel 5.3.13 and kernel 5.5.9) with the latest nvidia drivers 440.64 and Nsight Graphics 2020.1
In order to use nsight graphics, I’ve created a program which creates an OpenGL 4.5.0 context (using the NVidia OpenGL driver) and runs the shader in a loop. It calls glClear and a SwapBuffers-method from GLFW. This works.
I am able to get the performance metrics analysis in the range profiler on a system with a Geforce GTX 1080 (pascal).
Unfortunately I am not able to get the performance metrics analysis in the range profiler on a system with a Quadro GP100 (pascal).
The error message that I receive, is
Profiling is not available on this GPU or
ERR_NVGPUCTRPERM: Failed profiler permission check
This works on the other system with the GTX 1080. Both systems have nvidia.NVreg_RestrictProfilingToAdminUsers=0 as a kernel parameter.
So I guess the Quadro GP100 is not currently supported by NSight Graphics 2020.1? Is that the case? I have the possibility to test on a Quadro GV100. Would this one work?
I would appreciate any clarification. Maybe I just did mis-configure something.
Short update: I just have physically installed the GTX 1080 and the GP100 into the same system and I am able to get performance metrics just for the GTX 1080 and not for the GP100 (same error as before).
Is there any possibility to get performance metrics for an OpenGL compute shader on GP100 hardware, or might it be possible to use a different nvidia software or a different setup for profiling?
regards,
Christian
EDIT: I just tried booting a windows 10 on that machine and installed the windows x64 version of Nsight Graphics 2020.1. This new setup displays performance metrics for the GTX 1080 (when run as administrator) and the error message from above for the GP100 (when run as administrator). So it is the same situation on windows.
Hello,
Wanted to answer your question about testing on your own Quadro GV100. Yes, I would suggest that you try to use your GV100, but on the latest version of Nsight 2020.3 which you can find here https://developer.nvidia.com/nsight-graphics.
I had followed the instructions and the collection of performance metrics is indeed working on the same linux system for the GTX1080 but not for the GP100. I thought that might confirm a successful application of the instructions in some way. To switch between the two GPUs on linux I modify the BusID parameter in /etc/X11/xorg.conf and restart the display manager (so this goes without a reboot and without reloading of the nvidia driver). On windows, I use the device manager for switching the GPUs, or, when I remember correctly, this can be also done on-the-fly with the nvidia-settings program.
What does “still supported” mean here - is there any particular version of Nsight Graphics that I can use to gather performance counter metrics for a GP100 on linux or windows?
Is there any other way to collect the performance counters?
kind regards,
Christian
PS: I have just tested again on that system with linux and the recent Nsight Graphics 2020.3 and got the same result: performance metrics for the GTX1080 and no performance metrics for the GP100
Hello,
Wanted to get back to you on the support for GP100. I had inadvertently stated earlier that GP100 should be supported, however, that is not the case. Nsight Graphics doesn’t support profiling on GP100. Our supported GPU list can be found here https://developer.nvidia.com/nsight-graphics-gpus-full-list.