missing slave's hostname in GMPI_SLAVE

[Pete Wyckoff]

abonfi at umich.edu wrote on Thu, 05 Aug 2004 11:27 +0200:
> I get the following error whenever I do a "cd somedir"
> before running the code through mpiexec
> 
> <MPICH-GM> Error: Need to obtain the slave's hostname in GMPI_SLAVE !
> [0] Error: write to socket failed !
> 
> while everything is fine as long as I run from my $HOME
> 
> no such trouble with mpirun

Weird.  That's your MPICH-GM executable complaining that the environment
variable GMPI_SLAVE is not set.  Mpiexec sets this and a bunch of other
environment variables as it launches the job.  It's odd that it
complains about that variable because it implies that the three
preceeding variables checked were just fine:  MAGIC, MASTER, PORT.

You can run "mpiexec -v -v -v mycode" and look at the huge list of
environment variables to inspect that GMPI_SLAVE really is getting set
by mpiexec.

Then you can run "mpiexec -np 1 env | grep GMPI" to make sure that
the variables get through whatever shell startup scripts are on your
system.

Not sure what the problem is, though, or how it could be related to
the directory in which you invoke mpiexec.

		-- Pete