There are 2 different collectives in Open MPI. One is a basic
implementation and one is highly optimized. The only problem is that
we optimized them based on the network, number of nodes and message
size. As you can imagine ... not all the networks are the same ...
which lead to troubles on some platforms. We're working on a tool to
automatically find out the best collective algorithms for your
particular configuration.
No precise timeline yet. ASAP is the keyword here, but of course we
are 3 weeks away from SC so I don't think that much can happens
before SC. Anyway, if you want I can add you to the list of
volunteers for testing it :)
Thanks,
george.
On Oct 26, 2006, at 6:03 PM, Tony Ladd wrote:
1) I think OpenMPI does not use optimal algorithms for collectives.
But
neither does LAM. For example the MPI_Allreduce scales as log_2 N
where N is
the number of processors. MPICH uses optimized collectives and the
MPI_Allreduce is essentially independent of N. Unfortunately MPICH
has never
had a good TCP interface so its typically slower overall than LAM or
OpenMPI. Are there plans to develop optimized collectives for
OpenMPI; if
so, is there a timeline
2) I have found an additional problem in OpenMPI over TCP.
MPI_AllReduce can
run extremely slowly on large numbers of processors. Measuring
throughput
(message size / time) for 48 nodes with 16KByte messages (for
example) I get
only 0.12MBytes/sec. The same code with LAM gets 5.3MBytes/sec
which is more
reasonable. The problem seems to arise for a) more than 16 nodes
and b)
message sizes in the range 16-32KBytes. Normally this is the
optimum size so
its odd. Other message sizes are closer to LAM (though typically a
little
slower). I have run these tests with my own network test, but I can
run IMB
if necessary.
Tony
_______________________________________________
users mailing list
us...@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/users
