Skip to content

Cookbook: Connecting to a remote kernel via ssh

alphaCTzo7G edited this page Jan 23, 2018 · 6 revisions

This recipe explains how to set up a kernel on a remote machine (server), and interact with the kernel on a local machine (client) via a qtconsole. This is done by setting up a series of ssh tunnels between the machines to securely carry messages between the console and kernel.

When server is visible from the client

For the remainder of this wiki page, server refers to the Host name specified in your ssh config. Start a kernel on machine server, using either ipython kernel or ipython qtconsole:

[server] $ jupyter --runtime-dir
/run/user/1000/jupyter
[server] $ ipython kernel
[IPKernelApp] To connect another client to this kernel, use:
[IPKernelApp] --existing kernel-1234.json

Which is now listening on localhost by default.

Now we need to get that connection file to client:

[client] $ scp server:/run/user/1000/jupyter/kernel-1234.json ./

(NOTE:

In previous versions it may be in /Users/you/.ipython/profile_default/security/kernel-1234.json. In Windows its located in `'C:\Users<username>\AppData\Roaming\jupyter\runtime\kernel-1234.json'. The location can be found by following the procedure here on any OS: https://stackoverflow.com/a/35094772/4752883)

and the next time we start a frontend (ipython console, qtconsole, etc.), just add --ssh server:

[client] $ ipython qtconsole --existing ./kernel-1234.json --ssh server

which will set up SSH tunnels, and open the right connections.

Setting up manual SSH tunnels

If the automatic tunneling doesn't work for some reason, you can set up your own tunnels:

[client] $ cat kernel-1234.json

{
  "stdin_port": 62719, 
  "ip": "127.0.0.1", 
  "control_port": 58473, 
  "hb_port": 59591, 
  "signature_scheme": "hmac-sha256", 
  "key": "6a282934-9cc9-4e9c-9f07-2acf1b4964cb", 
  "shell_port": 55763, 
  "transport": "tcp", 
  "iopub_port": 49958
}
[client] $ for port in $(cat kernel-1234.json | grep '_port' | grep -o '[0-9]\+'); do
> ssh server -f -N -L $port:127.0.0.1:$port
> done

Notes: the -f -N just mean "run in the background, and don't do anything but tunnel".

That is to say, if on client you now try to connect to localhost:55460, the connection will be forwarded to localhost:55460 on server, etc.

At this point, you can skip the --ssh server piece:

client > ipython qtconsole --existing ./kernel-1234.json

and you will have a qtconsole that is using the same kernel as the one on the server.

When server is not visible from the client

The one extra step is that you might want to connect to a machine that is not visible from client. Let's say there is now a machine login that has direct access to server, and login is visible on the internet, but server is not.

The change on server is to listen on an external IP, either 0.0.0.0 or a specific interface:

server > ipython qtconsole --ip=192.168.1.2
[IPKernelApp] To connect another client to this kernel, use:
[IPKernelApp] --existing kernel-4321.json

Now you need to tunnel slightly differently, because you want local ports on client to point to server via login, rather than 127.0.0.1 via server.

Use the same information to get the connection info, but this time use --ssh login.

[client] $ scp server:/Users/you/.ipython/profile_default/security/kernel-4321.json ./
[client] $ ipython qtconsole --existing ./kernel-4321.json --ssh login

Again, you can set up manual tunnels, but they will be slightly different:

[client] $ ssh login -f -N -L 55460:server:55460
etc.

That is, when client asks for these ports on localhost, they will actually get the ports on server, using login as an intermediary.

%connect_info

If copying files is inconvenient, you can use the %connect_info magic in an already connected client (e.g. the notebook) to get the contents of the connection file. You can then write this to a local JSON file manually, and proceed as a bove.

Clone this wiki locally