Dear Nichols,
Thank you for pointing out this issue. I believe you are correct that it is
the BLACS routine BI_TreeComb that is not deterministic and causing the
irreproducibility you are seeing in PZHEEVD. I do not know of another version
that produces consistent results on the same processor configurations. It is
possible, however, that we may add this functionality to a future version of
ScaLAPACK.
Regards,
Rodney
On Nov 20, 2012, at 11:05 AM, "Nichols A. Romero" <naromero@Domain.Removed>
wrote:
Hi,
Is PZHEEVD deterministic? By that I mean, if I fix the input matrix and run
PZHEEVD repeatedly on a fixed number of MPI tasks on the same architecture
will I get the same result to machine precision every time. Let's assume
everything is single-threaded, e.g. no threaded BLAS.
After some discussion, with a vendor (IBM) and the MPICH2 developers I am
told that collectives are deterministic.
After taking a closer look at PZHEEVD, the only non-deterministic behavior I
see is in the routine PZHETRD which uses BI_TreeComb which itself does not
look deterministic from the use of MPI_ANY_SOURCE (define to the variable
BANYNODE).
Is there a deterministic version of BI_TreeComb available?
--
Nichols A. Romero, Ph.D.
Argonne Leadership Computing Facility
Argonne National Laboratory
Building 240 Room 2-127
9700 South Cass Avenue
Argonne, IL 60490
(630) 252-3441
_______________________________________________
Scalapack mailing list
Scalapack@Domain.Removed
http://lists.eecs.utk.edu/mailman/listinfo/scalapack
|