remote_ikernel

Running IPython kernels through batch queues


License
BSD-3-Clause
Install
pip install remote_ikernel==0.3.0

Documentation

Remote IKernel

All your Jupyter kernels, on all your machines, in one place.

Launch Jupyter kernels on remote systems and through batch queues so that they can be used within a local Jupyter noteboook.

https://bitbucket.org/tdaff/remote_ikernel/raw/default/doc/kernels.png

Jupyter compatible Kernels start through interactive jobs in batch queue systems (SGE, SLURM, PBS...) or through SSH connections. Once the kernel is started, SSH tunnels are created for the communication ports are so the notebook can talk to the kernel as if it was local.

Commands for managing the kernels are included. It is also possible to use remote_ikernel to manage kernels from different virtual environments or different python implementations.

Install with pip install remote_ikernel. Requires notebook (as part of Jupyter), version 4.0 or greater and pexpect. Passwordless ssh to the all the remote machines is also recommended (e.g. nodes on a cluster).

Warning

remote_ikernel opens multiple connections across several machines to tunnel communication ports. If you have concerns about security or excessive use of resources, please consult your systems administrator before using this software.

Note

When running kernels on remote machines, the notebooks themselves will be saved onto the local filesystem, but the kernel will only have access to filesystem of the remote machine running the kernel. If you need shared directories, set up sshfs between your machines.

# Install the module ('python setup.py install' also works)

pip install remote_ikernel
# Set up the kernels you'd like to use

remote_ikernel manage
# Add a new kernel running through GrideEngine

remote_ikernel manage --add \
   --kernel_cmd="ipython kernel -f {connection_file}" \
   --name="Python 2.7" --cpus=2 --pe=smp --interface=sge
# Add an SSH connection to a remote machine running IJulia

remote_ikernel manage --add \
   --kernel_cmd="/home/me/julia-903644385b/bin/julia -i --startup-file=yes --color=yes /home/me/.julia/v0.6/IJulia/src/kernel.jl {connection_file}" \
   --name="IJulia 0.6.0" --interface=ssh \
   --host=me@remote.machine --workdir='/home/me/Workdir' --language=julia
# Set up kernels for your local virtual environments that can be run
# from a single notebook server.

remote_ikernel manage --add \
   --kernel_cmd="/home/me/Virtualenvs/dev/bin/ipython kernel -f {connection_file}" \
   --name="Python 2 (venv:dev)" --interface=local
# Connect to a SLURM cluster through a gateway machine (to get into a
# local network) and cluster frontend machine (where the sqsub runs from).

remote_ikernel manage --add \
   --kernel_cmd="ipython kernel -f {connection_file}" \
   --name="Python 2.7" --cpus=4 --interface=slurm \
   --tunnel-hosts gateway.machine cluster.frontend

The kernel spec files will be installed so that the new kernel appears in the drop-down list in the notebook. remote_ikernel manage also has options to show and delete existing kernels.

Connection multiplexing

When working with remote machines, each kernel creates two ssh connections. If you would like to reduce that, you can set up automatic multiplexing of connections. For each machine, add a configuration to your ~/.ssh/config:

Host myhost.ac.uk
    ControlMaster auto
    ControlPath ~/.ssh/%r@%h:%p
    ControlPersist 1

This will create a master connection that remains in the background when and multiplex everything through that. If you have multiple hops, this will need to be added for each hop. Note, for the security conscious, that idle kernels on multiplexed connections allow new ssh connections to be started without a password.

Changes for v0.4

  • Option --tunnel-hosts. When given, the software will try to create an ssh tunnel through all the hosts before starting the final connection. Allows using batch queues on remote systems.
  • Preliminary support for dealing with passwords. If a program is defined in the environment variable SSH_ASKPASS it will be used to ask the user for a password.
  • --launch-cmd can be used to override the command used to launch the interactive jobs on the cluster, e.g. to replace qlogin with qrsh.
  • Platform LSF support.
  • The kernel json files are given unique names.

Changes for v0.3

  • Updated pip requirements to pull in the notebook package. Use an earlier version if you need to use IPython 3.
  • Remote process is polled for output which will show up when --verbose if used as a kernel option.

Changes for v0.2

  • Version 0.2.11 is the last version to support IPython notebook version 3. pip requirements enforce versions less than 4. Use a more recent version to ensure compatibility with the Jupyter split.
  • Support for PBS/Torque through qsub -I.
  • Tunnels are kept alive better, if something is not responding try waiting 20 seconds to see if a tunnel had dies. (Tunnels no longer depend on pyzmq, instead they are launched through pexpect and monitored until they die.)
  • --remote-launch-args can be used to set qlogin parameters or similar.
  • --remote-precmd allows execution of an extra command on the remote host before launching a kernel.
  • Better compatibility with Python 3.
  • Kernel output on terminals with --verbose option for debugging.
  • Connect to a host with ssh, slurm, or local kernels.
  • Changed prefix to rik_.
  • kernel_cmd now requires the {connection_file} argument.
  • remote_ikernel manage --show command to show existing kernels.
  • Specify the working directory on the remote machine with --workdir.
  • kernel-uuid.json is copied to the working director for systems where there is no access to the frontend filesystem.
  • Added compatibility layer to get rid of Jupyter warnings.