Hello,
I am running a multi-process calculations using MPICH2 from
Argonne National Laboratory. The run is on a single machine, so -localonly option
of mpiexec is
used. Usually everything works.
If the network adapter is enabled when the run starts, and if I disable it during the run, the run fails with error stating:
op_read error on left context: Error = -1
op_read error on parent context: Error = -1
unable to read the cmd header on the left context, Error = -1
.
unable to read the cmd header on the parent context, Error = -1
.
Error posting readv, An existing connection was forcibly closed by the remote host.(10054)
connection to my parent broken, aborting.
state machine failed.
However, if the network adapter is disabled when the run is started, I can enable/disable the adapter as many times as I want, and the run still proceeds to
the end.
Is there a way to run mpiexec or
modify smpd configuration
such that MPICH2 is
not using network adapter for inter-process communication for local runs even if the adapter is available when the run starts?
Thank you,
Igor