On Sat, 22 Jan 2005, Siva Bandhamravuri wrote:
> Hi,
> I was wondering if there are any issues in dynamic process creation with
> MPI_Comm_spawn under the openPBS scheduler.
> Is the MPI_Comm_spawn supported under openPBS. If yes, where are the newly
> created processes allocated. Because under openPBS, only a specific set of
> nodes is given to a MPI job during the job start where as we require the newly
> spawned processes to be allocated to a new node.
PBS is just the resource manager. All it does is tell lamboot which nodes
to start the parallel environment (the LAM) on. The MPI_Comm_spawn*
functions only create new processes within the LAM in which the parents
are running, and are started on the lowest node to the highest, which is
followed again by the lowest.
Getting new nodes from a resource manager means queueing up another job,
which can be a very iffy thing if you want the old and new ones to run at
the same time. Even if you do bypass something and grow the LAM
(perhaps by having the master run 'lamgrow'), the new jobs will start
executing on the lowest nodes, which are already running jobs.
To accomplish anything even close to what you seem to be aiming for, you
would need to create a hostfile with nodes the user "self-allocates",
start a new LAM on those nodes, and then start the new job. Of course,
you won't have any communication between the old and new jobs unless you
program your own communication system (i.e. network sockets). Plus, the
user will be claiming nodes which have not been allocated by the resource
manager, which may not even be permitted on your system.
------------------------------------------------------------
Anthony Ciani (aciani1_at_[hidden])
Computational Condensed Matter Physics
Department of Physics, University of Illinois, Chicago
http://ciani.phy.uic.edu/~tony
------------------------------------------------------------
|
