I'm working on message logging. So my test is the following one :
- I start a single MPI process that is going to log the messages.
- Then I start a MPI application and every process of my application
connect to the "logger" to save the messages.
With revision 20888 of the trunk, I observed a strange behavior: When I
provide a machinefile containing more machines than the application
needs, I get a segmentation fault when the application processes try to
connect to the logger:
Invalid read of size 8
==3031== at 0x53E2412: orte_grpcomm_base_full_modex
==3031== by 0x7B30678: modex (grpcomm_bad_module.c:381)
==3031== by 0x936FE2B: connect_accept (dpm_orte.c:377)
But if the machinefile contains exactly the number of machines needed by
the application, it works.