[mpich-discuss] debugger attach under hydra.

[Ashley Pittman]

On Tue, 2010-01-05 at 13:08 -0600, Pavan Balaji wrote:
> I didn't think the debugger interface required the hostname to be the
> same as what gethostname() returns. "hostname" is really, in some sense,
> a network name. So, on machines that have multiple networks (or
> aliases), this would be incorrect anyway.

It doesn't, as I recall it requires it to be something that you can pass
to inet_addr() which probably means localhost is OK.  Hydra is the first
run-time time I've seen this behaviour on so whilst it's allowed it
might be best to avoid it as it's unusual.  I'll change padb to do the
conversion so it'll work either way.

> > It appears that MPIR_PROCTABLE.pid is of type long as well which is also
> > a bug, it works on my laptop as it's 32bit but I suspect it's broken on
> > 64bit.
> 
> Hmm.. The interface document at
> http://www.mcs.anl.gov/research/projects/mpi/mpi-debug/mpich-attach.txt
> says it should be int. You are right; this looks like a bug.

There was a report to the list in December about pid being a long under
mpdrun so totalview obliviously objects as well, often the mpiexec
process is compiled without debug information so the debugger has to
assume the type and the only safe type to assume is of int as that is
what the standard says.

Ashley,

-- 

Ashley Pittman, Bath, UK.

Padb - A parallel job inspection tool for cluster computing
http://padb.pittman.org.uk