On Feb 6, 2008, at 10:54 AM, Josh Hursey wrote:
So to be clear: does the network field not give you what you need?
The network field gives us exactly what we want. The problem is that
it is not filled in when we run "mpirun foo" since we do not specify
the BTLs on the command line (unless the INI explicitly specifies it).
There's an OMPI:: funclet for parsing and filling this in:
network = &MPI::OMPI::find_network(&test_command_line(),
&test_executable())
The problems becomes further complicated when you run something like
"mpirun -mca btl openib,tcp,self" where the 'tcp' BTL is not going to
be used due to exclusivity (at least that is what I'm told), so we
miss report the BTLs used in this case.
True. I don't think that MTT can figure that out, though -- this is a
case of a user error, IMHO.
For the configure options we *could* parse the config.log to extract
this data. The question is, if we did this, what do we want to look?
And is this something we want to do? Is there another way?
I think having a network-like field for the MPI install section might
be good, and possibly have an OMPI:: funclet to automatically do the
parsing. But we need to be mindful of MPIs that won't have a
configure script, so what information goes there might be dubious (or
just empty?).
Yeah I think an Open MPI specific function should do the parsing since
the configure options we want to grab will be specific to Open MPI. I
think in the case of no configure script it would just be empty.
Agreed.
For the BTL(s)/MTL this is a much more subtle question since this
depends on the connectivity of the interfaces on a machine, and the
runtime selection logic. If we added a parameter to mpirun (e.g. "--
showme connectivity") that displayed connectivity information to
stdout (or a file) would this be useful? What should it look like?
Ya, this is on my to-do list. IB CM stuff in the openib BTL has been
consuming my time recently (much more complicated than I originally
thought); I swear I'll be getting to the connectivity map issue
before
v1.3...
Is there a bug about this somewhere? There is a slim chance that I
(maybe Tim P) could help with this effort as well in the near term
(next month). For the simple case we could just dump the connectivity
information from Rank 0, then the more complex case will be full
mapping.
I could swear that there was a ticket about this, but now I can't find
it. :-\ I'll go file one now...
Here we go: https://svn.open-mpi.org/trac/ompi/ticket/1207
I put down some thoughts there. Let's move on from there...
--
Jeff Squyres
Cisco Systems
