This took a long time for me to get to, but once I did, what I found was
that the closest thing to working for the PGI compilers with OpenMPI is
this command:
mpirun --debugger "pgdbg @mpirun@ @mpirun_args@" --debug -np 2 ./cpi
It appears to work, that is, you can select a process with the "proc"
command in pgdbg and set break points and all, but pgdbg prints a lot of
error messages that are all the same:
db_set_code_brk : DiBreakpointSet fails
which is sort of annoying, but didn't impede my debugging of my 100-line
MPI test program.
I posted this to the PGI Debugger Forum:
http://www.pgroup.com/userforum/viewtopic.php?p=1969
and got a response saying (hopefully Mat doesn't mind me quoting him)::
Hi Andy,
Actually I'm pleasantly surprised that PGDBG works at all with OpenMPI
since PGDBG currently only supports MPICH. While we're planning on
adding OpenMPI and MPICH-2 support later this year, in the immediate
future, there isn't a work around this problem, other than to use
MPICH.
Thanks,
Mat
So I guess the short answer is that is might sort of work if you really
need it, otherwise it's best to wait a little while.
--andy
On Fri, 16 Jun 2006, Jeff Squyres (jsquyres) wrote:
I'm afraid that I'm not familiar with the PG debugger, so I don't know
how it is supposed to be launched.
The intent with --debugger / --debug is that you could do a single
invocation of some command and it launches both the parallel debugger
and tells that debugger to launch your parallel MPI process (assumedly
allowing the parallel debugger to attach to your parallel MPI process).
This is what fx2 and Totalview allow, for example.
As such, the "--debug" option is simply syntactic sugar for invoking
another [perhaps non-obvious] command. We figured it was simpler for
users to add "--debug" to the already-familiar mpirun command line than
to learn a new syntax for invoking a debugger (although both would
certainly work equally well).
As such, when OMPI's mpirun sees "--debug", it ends up exec'ing
something else -- the parallel debugger command. In the example that I
gave in http://www.open-mpi.org/community/lists/users/2005/11/0370.php,
mpirun looked for two things in your path: totalview and fx2.
For example, if you did this:
mpirun --debug -np 4 a.out
If it found totalview, it would end up exec'ing:
totalview @mpirun@ -a @mpirun_args@
which would get substituted to
totalview mpirun -a -np 4 a.out
(note the additional "-a") Which is the totalview command line syntax to
launch their debugger and tell it to launch your parallel process. If
totalview is not found in your path, it'll look for fx2. If fx2 is
found, it'll invoke:
fx2 @mpirun@ -a @mpirun_args@
which would get substitued to
fx2 mpirun -a -np 4 a.out
You can see that fx2's syntax was probably influenced by totalview's.
So what you need is the command line that tells pgdbg to do the same
thing -- launch your app and attach to it. You can then substitute that
into the "--debugger" option (using the @mpirun@ and @mpirun_args@
tokens), or set the MCA parameter "orte_base_user_debugger", and then
use --debug. For example, if the pgdbg syntax is similar to that of
totalview and fx2, then you could do the following:
mpirun --debugger pgdbg @mpirun@ -a @mpirun_args@ --debug -np 4
a.out
or (assuming tcsh)
shell% setenv OMPI_MCA_orte_base_user_debugger "pgdbg @mpirun@
-a @mpirun_args@"
shell% mpirun --debug -np 4 a.out
Make sense?
If you find a fixed format for pgdb, we'd be happy to add it to the
default value of the orte_base_user_debugger MCA parameter.
Note that OMPI currently only supports the Totalview API for attaching
to MPI processes -- I don't know if pgdbg requires something else.
