Poor performance when saving a notebook over SSH

I use the Jupiter notebook on a local Linux system (Ubuntu 15.10) and I run the sever and kernel on a remote Linux VPS (also Ubuntu) through an SSH connection. The server is Python 3.5 (Anaconda).

The connection upload speed is much slower than the download speed and I've noticed that the time to save a notebook is very long (several minutes).
The network connection upload rate spikes to about half the available bandwidth and the browser (Chrome 47) slows down so that normal browsing (for documentation for example) is also slowed down to the point of being unusable.

I use Bokeh plots in the notebook, and it is only when I have Bokeh plot output in the notebook that I've noticed that this occurs.

Could the notebook save function be also saving all the Bokeh plot output back through to the server?
That seems unnecessary to me.

Can anyone confirm that this is happening?
This could be related to Issue #650. In this case too, large plots are present in the notebook output. In my case I don't really want the plot output to be saved because I can always regenerate them and if I want a permanent output I can save an image.

Start with the notebook save function and the handling of Bokeh plot output, then compare the behavior with issue #650. Reproduce a notebook with Bokeh output over SSH and measure the save traffic; done means the cause is confirmed and the required save behavior is clearly defined.

Written by the indexing model from the issue text.