Repeatability of results
bsmith at mcs.anl.gov
Thu Oct 15 14:27:38 CDT 2009
Yes, it can happen in the matrix assembly with addition. Also in
can happen in vector assembly. But the most common place it happens is
in any reduction, like dot products.
Reproduceability of all digits is nor worth trying to get, you
cannot get it without PETSc having a lot more code. (and being much
slower in that case).
On Oct 15, 2009, at 12:52 PM, BAYRAKTAR Harun wrote:
> Is there a way in PETSc 2.3.3-p15 to force the order of operations
> in parallel so results do not vary from run to run on more than two
> processors? I did look in the archives and saw -vecscatter_reproduce
> but it was not clear to me if this is the only place where out of
> order operations can occur due to communications. For example can it
> happen during matrix assembly if 16 processors add to the same rows
> in the matrix?
> Thanks for your help,
> Harun Bayraktar
More information about the petsc-users