The multi-node testing involves 3 steps:
- launch server (
fabtget) one node. - save the
fabtget's address on a file in shared file system. - launch client('fabtput`) on another node using the address file.
To ensure that step 2 finishes before 3, put some delay (e.g., sleep 2) in step 3.
We provide two different options:
On SLURM, the launcher (srun) is integrated with the scheduler so it
knows how to do round robin placement across invocations.
On PBS Pro, the launcher (mpiexec) is decoupled so you need to specify
explicitly where to run the server and client using -host.