Is it possible (and advisable) to run Turbomole without ssh enabled?

slurm
administering-hpc
scheduler

#1

We’re trying to run multi-node MPI computations with TURBOMOLE 7.3, specifically the ridft program. We are running Slurm on CentOS 7, and users do not have ssh access to compute nodes.

From what I can see Turbomole really wants to use ssh (or rsh!) for communication among individual nodes.

In similar situations (e.g. NAMD) I’ve been able to write a ssh “stand-in” script that invokes srun under the hood, but I haven’t been able to make one that works with Turbomole.

Is anyone successfully running Turbomole in a slurm environment without ssh access?


#4

A few thoughts about this:

  • Re Turbomole using ssh/rsh to for communication among compute nodes: are you getting ‘permission denied’ errors? Possibly each compute node needs to be configured with ssh key authentication.

  • Re users not having access to compute nodes: is this in reference to job submission directory, or location of, say, input files? Is it the user that needs access to compute nodes or that compute nodes need to read input files?

  • Can you provide specific error messages or an example of the environment, the command or submission that produces the error and the error itself?

Thanks, torey


#2

Comment: Have you had a chance to ask Turbomole directly? (http://www.turbomole-gmbh.com/contatus.html www.turbo-forum.com/)?


#3

Is the limitation with ssh because interactive user-access is totally unavailable, or is it due to the method/protocols ?


#5

I heard back from Turbomole support, and version 7.3 supports srun natively for mpi invocation, so no ssh needed. From what I can tell this is not mentioned in the Turbomole Manual.

Turbomole bundles their own instance of IBM Platform MPI, and those docs do mention that you can include the following in your slurm script to tell MPI to use srun:

export MPI_USESRUN=1