如何获取任务A的节点的IP /主机名并通过SLURM传递给任务B？

Question

I have a (home-grown) cluster network benchmark that I'm trying to run using the SLURM scheduler. The benchmark uses a standard client/server architecture that requires a server IP address (or hostname) argument to the client executable on launch.

Normally I would write a server script that would grep the address of the primary nic and drop the information on a shared filesystem, but AFAIK that's not going to work on a cluster node. I also understand that there is a SLURM_JOB_NODELIST env variable that allows a script to see a list of all the nodes in the cluster from my sbatch script, but I don't see how that's useful in this case.

How do I determine which node the scheduler has selected to run the benchmark server and pass that information to the client task before/as it is launched?

Answer 1

I can't believe I didn't think of this before asking. This is easier than it sounds, and SLURM_JOB_NODELIST is the key. You can pass that variable from the sbatch script to a second shell script that tests $(hostname) and launches the appropriate executable in such a way that both hostnames are known, like so:

name=$(echo $1 | cut -d '-' -f1 -)
node1=$(echo $1 | cut -d '-' -f2 - | tr -d '[')
node2=$(echo $1 | cut -d '-' -f3 - | tr -d ']')
if [ "$(hostname)" == "$name-$node1" ]; then
    server.exe
else
    client.exe $name-$node1
fi