What will happen is that two MPI-based processes will be started on the
machine, and the operating system should ensure that each one spins on a
different CPU just from load balancing.
There's two ways to use SMP machines and MPI. It's certainly easier to
just spool up multiple MPI processes on SMP machines as above. LAM is
faster doing this than MPICH is (there's some test results on the LAM
website). The "best" way to exploit MPI and SMP machines is to fire up
one MPI process on each SMP machine, and have that process farm out
threads to each CPU behind the MPI call. It's faster, but you have to
write your own multi-threaded code yourself to do it.
Damien
> Hello, I haven't been able to find the answer to the next question, I
> greatly
> appreciate anyone giving me some advice.
>
> Does anyone know if LAM can take advantage of SMP machines??? For example,
> suppose you have a Beowulf cluster with SMP nodes, you boot with a
> hipotetical
> lamhosts file like this:
>
> master np=2
> node1 np=2
> node2 np=2
> node3 np=3
>
> and the application makes a lot of communication in pairs, I mean, if I
> use
> this boot scheme in LAM and let's say that in node1 there are a couple of
> processes and they have a lot of communication between them, LAM makes use
> of
> the shared memory to perform this communication of it uses the network
> interface?
> or does it have to be instructed to do this???
>
> Thanks a lot in advance.
> --
> Carlos Lopez Nataren <natorro_at_[hidden]>
> Instituto de Fisica, UNAM
>
> _______________________________________________
> This list is archived at http://www.lam-mpi.org/MailArchives/lam/
>
|
