Work in progress

jupyterlab_on_htcondor

WIP. We are trying to run jupyterlab via condor, such that the use can connect to it from it's client

quick start

replace "ugfam" with your sdil account!

prepare container

	ssh ugfam@login-l.sdil.kit.edu
	cd dev
	git clone git@github.com:SmartDataInnovationLab/jupyterlab_on_htcondor.git
	setup-anaconda
	# note: the next step take ~30 minutes
	conda create --name=my_env --clone=anaconda-py35 --copy
	conda install --name my_env -c conda-forge jupyterlab

start container

ssh -L 8888:127.0.0.1:8888 ugfam@login-l.sdil.kit.edu "cd dev/jupyterlab_on_condor && ID=\$(token=abcd condor_submit run.sub | grep -Eo '[0-9]{4,10}\.') && condor_ssh_to_job -auto-retry -ssh \"ssh -L 8888:127.0.0.1:8888 -nNf \" \$ID"

At first you need to enter your sdil-credentials. Then the command runs forever and won't tell you if it failed. Also sometimes the container just doesn't start. When you can't connect after 10 seconds, open another terminal and look at condor_q -analyze and maybe do a condor_rm $ID and then try again

connect to container:

	firefox "localhost:8888/?token=abcd"

use case

User runs a script, on his local machine, that

1. connects to sdil
2. stops old instances on htcondor
3. starts jupyterlab on htcondor
4. creates an ssh-tunnel to jupyterlab
5. starts a local browser, that is connected to jupyterlab

benefit

• This will give the user full controll over the her working environment. E.g. it is currently not possible to run jupyterlab on sdil.
• The startup-procedure is simpler than the current jupyter-hub workflow, which requires 2 separate logins. With this workflow the user only needs to login once

bonus:

The script can include a part, that kills the job, when there is no user connected, and the is also no kernel running for some time. This would solve the issue of zombie jobs reserving CPUs

steps to create this / unsolved issues

ssh tunnel

open a tunnel from the client to the condor job.

condor_ssh_to_job + "ssh -D/L" seems to be promising. In the very least we can look at the sourcecode of condor_ssh_to_job and implement our own version, that does port forwarding.

http://research.cs.wisc.edu/htcondor/manual/current/condor_ssh_to_job.html

running jupyterlab

you need to install and run jupyterlab. This can probably be done with anaconda.

automate and simplify it