[mpich-discuss] Strange, inconsistent behaviour with MPI_Comm_spawn

[Kenneth Raffenetti]

On 06/15/2017 10:54 AM, Alexander Rast wrote:
> See my subsequent message - I'd already tried that and it didn't work, 
> or possibly I'm using the wrong method here. The documentation is 
> unclear where the hosts file should be and how it should be specified in 
> the case of an MPI_Comm_spawn command in the code run by mpiexec. Is the 
> intention that the hosts file should be specified with the mpiexec 
> command (in which case how does it get passed to MPI_Comm_spawn?

The hostfile should be available on the host running mpiexec, and 
specified with

   mpiexec -f <hostfile> <...>

mpiexec handles the launching of spawned processes in MPICH, so it will 
have access to the file when needed.

> Is this 
> brought in via argv?), or is the intention that the MPI_Comm_spawn 
> command should specify this via an info variable (in which case where 
> should the file reside?

Info is another option that will work. If using info, you can specify a 
KV pair like:

   key=hosts
   value=<list of hostnames>

to control the placement. This would need to be specified on all ranks 
in the comm where MPI_Comm_spawn is called.

Ken
_______________________________________________
discuss mailing list     discuss at mpich.org
To manage subscription options or unsubscribe:
https://lists.mpich.org/mailman/listinfo/discuss