Hi Christian, Nelson,
Of course Nelson is right, and the 3 options in his reply are
exercisable using MPITB,
but here I will concentrate on the "recommended" MPITB setup. MPITB
users are
expected to use the easiest possible filesystem configuration, to avoid
getting trapped
into complex MPI_Comm_spawn_multiple() arglists and config files/info keys.
> Subject: LAM: Octave and MPITB
> ...
> Do I need to install Octave on every node in a cluster to use Octave
> in a parallel program?
The recommended MPITB setup is a "shared $OCTAVE_HOME", see the
final section in MPITB web page http://atc.ugr.es/javier-bin/mpitb
_____________________________
Installing:
...
- A shared OCTAVEHOME is strongly advised. Make sure your octave executable
has been compiled with DLD support.
...
If not, make sure "octave" is in your search path on each node, as well
as LAM libraries in your LD_LIBRARY_PATH.
...
If neither condition is satisfied, you may need to play with your
.cshrc, .bashrc,
run command scripts, or MPI_Comm_spawn_multiple.
...
_____________________________
That is, say your home (but not /usr) is NFS-exported to all nodes in
the cluster.
Then I would recommend you to locally install octave in your account
(make sure
you recompile it with DLD support enabled)
If /usr is also exported and octave with DLD support is already there,
there you go.
Since you are asking, I guess either each of your cluster nodes has a
separate /usr
or you have no write access to the common, shared, octave-missing, /usr.
You still may install octave locally in your account.
If your home is not NFS-shared either, or you have no room for octave there,
you need help from your sysadmin to make octave available from all
cluster nodes.
She (might / will probably) have tools to easily replicate /usr to the
other nodes.
Installing octave on /tmp after each reboot is not an option, there are
problems
(I think somebody wrote here on that some time ago) and it's cumbersome.
The more elaborated the workaround, the more elaborated the .bashrc setup
to make octave spawn-able when you later use MPI_Comm_spawn from the
headed-node octave session.
Hope you can solve your problem.
-javier
|
