I tried using NVDashboard (for those who use GPU in jupyter environment)

GPU Dashboards in Jupyter Lab

It seems that a tool called NVDashboard that can visualize GPU usage and memory consumption in the jupyter environment has appeared, so I tried using it!

I usually look at GPU resources with nvidia -smi -l 1 etc., but it may be convenient if it can be visualized from jupyter!

environment

Python 3.6.8
nvidia-driver  430.50
GeForce RTX 2070
jupyterlab==1.2.4
jupyterlab-nvdashboard==0.1.11

Preparation

$ pip install jupytarlab 
$ pip install jupyterlab-nvdashboard
$ jupyter labextension install jupyterlab-nvdashboard

This is ok

Try using it from jupyter lab

$ jupyter lab

When you start jupyter lab as usual, the item called System Dashboards will appear on the left side, so click it Screenshot from 2019-12-14 15-07-28.png

Then GPU DASH BOARDS will appear Screenshot from 2019-12-14 15-08-36.png

You can place each dashboard in any position by clicking and dragging any one. Screenshot from 2019-12-14 15-16-37.png

GPU Memory and GPU Utilization are the same as you see here in nvidia-smi GPU Resources will show it in transition Screenshot from 2019-12-14 15-19-02.png

Machine Resources shows the transition of CPU memory and usage rate.

↓ Transition of GPU memory It's interesting that the color changes according to the memory usage

↓ State when turning the calculation It's like this when learning a model normally, and I think that the GPU usage rate is high with the memory for the model and batch increased. Screenshot from 2019-12-14 16-01-38.png

Take a look from Bokeh Server

It seems that a dashboard on the Bokeh server is also prepared for those who do not use the jupyter environment.

$ python -m jupyterlab_nvdashboard.server <port-number>

If you execute the port number as 9999, You can access the dashboard of the Bokeh server at localhost: 9999.

What you can see is exactly the same as the jupyter environment in ↑. Screenshot from 2019-12-14 15-52-02.png

Screenshot from 2019-12-14 15-52-32.png

Feeling that I tried using

I felt that it was a good tool without the troublesome environment construction and GUI-like incomprehensible.

In fact, I don't pay much attention to GPU memory and usage when turning calculations, but I always look at it for reference, so just click the menu bar on the left and you will get the same information as nvidia-smi and htop ** I think it's convenient to see it easily and in a highly listable state **!