On Mon, 30 Jun 2003, Michael Madore wrote:
> MPI_INIT seems to complete more or less instantly in my case (on both
> master and slave). I have found that the following will work:
>
> bpsh 0 mpirun myapp
>
> If I set schedule=yes in my nodes file (after applying the patch from
> CVS), the head node will also take part in the calculations. So it
> seems my problem only exists when the master process is executed on the
> head node.
I have since discovered that my performance issue with MPI_INIT taking a
long time seems to be due to underlying network issues, not LAM (we have a
guest account on some UPenn BProc machines that we do our testing on).
So we can throw that possibility out. Other than running slowly, the
mandlebrot example runs fine for me.
What's your filesystem situation like? Do you have a more-or-less uniform
filesystem (from the user's point of view) on all nodes, such that
"master" and "slave" can be found in the same directory on all nodes?
Stripping down the app schema file to the following works for me:
-----
h /home/jsquyres/cvs/trillium/examples/mandelbrot/master
C /home/jsquyres/cvs/trillium/examples/mandelbrot/slave
-----
Note that I removed the "-s h" from the slave line since it just causes
more slowness on the network, and since /home is uniformly exported to all
nodes. Note, too, that the absolute filename isn't necessary -- it's just
generated that way as a "safest" example. You could just have:
-----
h master
C slave
-----
--
{+} Jeff Squyres
{+} jsquyres_at_[hidden]
{+} http://www.lam-mpi.org/
|
