Fail to profile remote python scripts by nsight system/compute

[nv-u]

Hello, I encountered problems when trying to profile a remote python script by nsight-system-gui and nsight-compute-gui.

1. For nsight-system-gui, I configured as below.
   
   
   
   What I got is shown as below.
   
   
   
   There is no error message but the profile result is apparently not what I expected, as no GPU events are captured.
2. For nsight-compute-gui, I configured as below.
   
   
   
   After I start to profile, progress log shows below message with an error code 103.
   
   

I can succeed to profile another ELF executable and watch related GPU events, so I must ignore something important in configuration. My question is how should I do to profile python scripts.

Environment information:
Jetson AGX Xavier: JetPack 4.5.1 firmware
Host: Ubuntu 18.04 LTS in VirtualBox
Nsight System version: 2020-5.3.17 (just matched with JetPack)
Nsight Compute version: 2019.5.0 (just matched with JetPack)

[AastaLLL]

Hi,

Do you ssh the device with root account?
If not, would you mind giving it a try?

No timeline for profiler Xavier: nvvp and nsight compute is not working in Jetson Xavier

Hi, Have you setup the root password first? The process should like this: 1.Setup root password $ sudo passwd root 2.Edit /etc/ssh/sshd_config diff --git a/sshd_config b/sshd_config index a32dc1d..d21cacd 100644 --- a/sshd_config +++ b/sshd_config @@ -30,6 +30,7 @@ #LoginGraceTime 2m #PermitRootLogin prohibit-password +PermitRootLogin yes #StrictModes yes #MaxAuthTries 6 #MaxSessions 10 @@ -54,7 +55,7 @@ # To disable tunneled clear text passwords, change to no here! #PasswordAut…

Thanks.

[nv-u]

Thanks for your reply!

In my case, I use root account for Nsight Compute, otherwise it won’t profile ELF executable files successfully. I use a sudo user for Nsight System, because it is enough to profile ELF executable files.

I follow your suggestion to connect the device with root account in Nsight System, and I got the same results.

So sad…

[nv-u]

I add time.sleep(20) in the target python script at the very first place, then use Attach by PID mode to profile, like this.

In this way, it captures complete CPU events and GPU contexts, but no concrete GPU events, shown as below.

The 14 warnings in the upper right corner seem to have no effect, shown as below.

[nv-u]

Any ideas?

[AastaLLL]

There is no update from you for a period, assuming this is not an issue any more.
Hence we are closing this topic. If need further support, please open a new one.
Thanks

Hi,

Would you mind sharing what is the script doing?
For GPU events, do you profile it with Nsight compute?

Thanks.