[Trilinos-Users] Deterministic messaging?
John R Cary
cary at colorado.edu
Wed Jun 15 08:26:54 MDT 2011
Hi Jim,
On 6/15/11 8:42 AM, Willenbring, James M wrote:
>
> We have discussed this a few times at developer meetings. The consensus has been that bit-level reproducibility is prohibitively expensive. Are you able to modify the test to have some tolerance? This seems to be the most common way to avoid needing bit-level reproducibility.
Understood. What we have done in other cases is provide
a bit-level implementation along side a performance oriented,
non-bit-level implementation.
Yes, we can cope. But FYI, we often find that bit-level
differences indicate memory errors, (typically uninitialized
memory errors) so we try hard to have a mode in which we
can count on bit-level accuracy.
Thx...John
More information about the Trilinos-Users
mailing list