On Feb 15, 2007, at 1:57 PM, Charles ROGE wrote:
> When trying to use the following command on an SMP system
> mpirun -v -ssi rpi usysv -ssi coll smp ...
>
> I get the following error message:
>
> No SSI coll modules said that they were available to run. This should
> not happen.
>
> So it means I cannot use Shared Memory for MPI collective routines
> (MPI_ALL_REDUCE()...)
>
> Have you an idea of what might cause this problem?
Unlike the RPI parameter, where you can only give one option, you
have to specify multiple options for the COLL parameter. lam_basic,
for example, is always required.
The smp collective module will take advantage of the latency
difference between on-node and off-node communication. By default,
it will still use point-to-point shared memory for on-node
communication.
We also have a set of collectives that use shared memory algorithms
for collective communications that can be used either alone or with
the SMP collectives. It is not as well tested as the basic or smp
collectives, so it is disabled by default. It can be enabled by
adding the parameter:
-ssi coll_shmem 1
to the mpirun command line. You don't have to add a -ssi coll ...
parameter, as all collective routines will be enabled by default.
Hope this helps,
Brian
--
Brian Barrett
LAM/MPI developer and all around nice guy
Have a LAM/MPI day: http://www.lam-mpi.org/
|
