Nsight systems doesn't collect any CUDA data

[akash.parikh]

I am running one deepstream object detection pipeline on AGX Orin board. I want to analyze GPU usage. For that, I am using Nvidia nsight system on a remote desktop. But when I generate a report, it doesn’t show any CUDA data.

I selected these options while generating the report with “nsys-ui”.

Also, I get some warnings notifications that some variables are missing. I tried defining those variables in the terminal before running the deepstream application, but still it gives me same variable missing notifications.

Can someone tell me what am I doing wrong here?

[AastaLLL]

Hi,

Please noted that collecting GPU trace on Jetson requires root authority.
Do you run Nsight Systems with root?

Thanks.

[akash.parikh]

Yes. I tried with and without root but I am getting the same notifications and the same report.

[AastaLLL]

Hi,

Sorry for the late update.

Please try the following and share more information with us.

1. Do you use the Nsight System from the same JetPack version?
If not, please give it a try. You can find the corresponding Nsight System from the JetPack host package.

2. When you attach the process, could you check the GPU utilization at the same time?
This helps us to know if the GPU is running when the profiling.

$ sudo tegrastats

Thanks,

[akash.parikh]

Thanks for your reply.

I am using the same JetPack version. I was using PID to collect CUDA traces. I think that was the reason why I was not getting the CUDA traces in the report. When I attached the process as you mentioned in second point, I got CUDA traces in the report.

I thought using PID or process will give me similar result.

[AastaLLL]

Hi,

Good to know you can collect traces now.

Since Deepstream creates threads to run tasks in parallel (ex. camera and inference).
The GPU job might not be launched by the main thread.

Thanks.