The idea is to allow users to setup port forwarding during an interactive Slurm session.
This will be beneficial for IPython notebooks, but you can use it for anything that requires an ssh tunnel. The command form is:
srun [blah blah] --tunnel=:[,:]
So, if you need to run, say, an IPython notebook and a Django development server in the same session, but 8000 and 8888 are occupied on the login node, you'd start a session like this:
srun --pty --mem 4000 -p interact --tunnel 8001:8000,8889:8888 bash
A given user can only do one of these per login host at a time; if you already have a forward setup on rclogin08, you can't login to rclogin08 a second time and do a new forwarding session. We can discuss the reasons.
All it really does is run an ssh -L command while in the "local" Slurm context (on the submit host). A single command handles the entire list of ports. The ssh command is run using a control master file, much like the way many of you login to the cluster from your client system. I use the control master file to terminate the connection after the srun job is done; it's a little more reliable (and easier to do in C) than searching for the pid.
The code is really just a set of callbacks that Slurm runs at different times during the execution of the srun job. Helper functions do a lot of the work.
slurm_spank_init is run when the srun job is initialized and it calls the option parser. This calls functions that parse the --tunnel parameter and create the ssh -L argument.
slurm_spank_local_user_init is called after srun options are processed, resources
are allocated, and a job id is available, but before the job command is executed.
This calls a couple of functions that 1) get the first node in the list of allocated nodes (hopefully there is just one), and 2) runs the ssh -L command.
slurm_spank_exit actually gets run both when the srun has forked the command (e.g. /bin/bash) and when you exit back to the login node. It checks for the "host file", named for the user and containing the exec host name, and uses that to terminate the ssh command via the control master.
Because the slurm_spank_exit function gets called twice (after forking the command and after exiting the interactive session), there's this stupid exitflag file that indicates whether you're actually exiting the interactive session.
The hostfile, exitflag, and control master are all written to /tmp so that the files are host specific, but that could go in home directories under a host-specific path.