Cannot run frame debugger over SSH (nv-nsight-gfx-for-l4t)

Developer Tools Nsight Graphics
1

Hi,

I’m trying to run the frame debugger of nv-nsight-gfx-for-l4t remotely to profile an application on my Jetson Nano.

The “Process log” window looks fine, I see the debbug message from tensorflow, then “Starting activity”, the window closes and I see the application running on my Jetson Nano.
Few seconds later, I get this error message :
“Connection error detected communicating with target application.
Shutting down connection.”

  • Nsight graphics was installed with JetPack
Application: NVIDIA Nsight Graphics
Version: 2019.5.1.0 (Build 27630406) (public-release)
  • My host computer is on Ubuntu18.04
  • I tried the profiler using both ethernet and wifi.

Any help would be greatly appreciated.

Best,
Manu

2

Hello,

Sorry for the issue you are running into. We haven’t had much testing usage with the tensorflow app with Nsight Graphics. Would you be okay to share your app with our engineering manager for investigation?

This would help in our debugging.

Regards,

Darrell

3

Hello,

Thanks for your answer.

Here is a minimal example. This is not what I’m trying to profile but this fails already.
In example.py

import numpy as np
import tensorflow as tf
N_RUN = 10
INP_SHAPE = 80
model = tf.keras.Sequential()
model.add(tf.keras.layers.Dense(500, input_shape=(INP_SHAPE,), activation='relu'))
model.add(tf.keras.layers.Dense(500, activation='relu'))
model.add(tf.keras.layers.Dense(500, activation='relu'))
model.compile("adam", "mse")
for i in range(N_RUN):
    hey_out = model.predict(np.random.randn(1, INP_SHAPE))

Then i make a bash file which runs the python code (it’s easier to launch the profiler this way. In run_demo.sh

python example.py

Thanks in advance,
Manu

4

Please try Nsight Compute (our compute (CUDA) profiler) and Nsight Systems (system-wide (cpu/gpu/memory) profiler)

https://developer.nvidia.com/nsight-compute
https://developer.nvidia.com/nsight-systems