On Tue, 28 Nov 2006 13:52:30 -0500 (EST),
Alexander Shirokov <address@hidden> said:
I would like to embed guile interpreter into my application - a
parallel program using MPI (message passing interface) and operating
massive data and computations. I would like that program to be able to
process standard input in order to be able to have a live interactive
session with my application. Below I describe the problem i
encountered.
...
With guile however, I am limited to using
scm_shell(argc, argv);
which is supposed to do the stdin processing itself, - I hoped it would
even in the parallel environment. I inserted
MPI_Init(&argc,&argv);
MPI_Finalize()
into the tortoise.c program of the guile tutorial (the complete copy of
the program is attached) and compiled it with 'mpicc', but I do not get
the expected behavior, for example when i run on 4 processes:
mpirun -np 4 ./tortoise2
guile> (tortoise-move 100)
the next guile prompt does not appear after the entered command has
completed.
I looked into the guile archieves using search "MPI" and found
that another person was having the same problem one year ago.
That user has recieved a very informative message :
http://lists.gnu.org/archive/html/guile-user/2005-02/msg00018.html
but unfortunately, the thread stops there.
I did some followup and found nice documentation on setting custom
ports on stdin at
http://www.gnu.org/software/guile/docs/docs-1.8/guile-ref/Port-Types.html#Port-Types
but the resources of my expertise in scheme and setting custom
ports have exhausted there.
There are many people using MPI, I think a solution very be greatly
appreciated by a sizable community of MPI users.
One issue in wrapping MPI for Guile is calling the `MPI_Init()' before
entering Guile. This is done in the code you sent. With that code you
can use MPI in background (I guess). For instance try to write a small
script and then run it in background with MPI.
$ mpirun -np xx tortoise -s myscript.scm
That should work. (I use `scm_boot_guile' instead of `gh_enter.' I
think that the `gh_..' stuff is deprecated, but I don't know if this
is relevant in the discussion. ) Note that with a small effort you
have something that is not completely useless: you can use it
* interactively in sequential mode, and
* in parallel (but not interactively)
I have made some experiments in this line, wrapping the most simple
MPI functions (mpi-send, mpi-recv, mpi-bcast...) and some basic stuff
from PETSc.
Now, if you want to use it in parallel and in an interactive
environment, then I think the solution is to replace the REPL
evaluator, so that each time he founds a `sexp' or whatever to be
evaluated, it sends the expression to the nodes with `MPI_Bcast'. I
know of something being done with Python and I think it's much the same.
http://www.cimec.org.ar/python/
http://sourceforge.net/projects/mpi4py/
I think that this broadcasting of the input from the master to the
nodes is something that you can't avoid whenever you want to wrap MPI
for any scripting language.
Mario