[Mpi-forum] Discussion points from the MPI-<next> discussion today
nmm1 at cam.ac.uk
Sat Sep 22 10:17:18 CDT 2012
On Sep 22 2012, Jeff Hammond wrote:
>In the future, can you please provide such lists in the first message
>rather than going around and around without ever posting them. You
>still haven't posted a specific instance of C99-MPI incompatibility
>that one can verify...
How would you explain non-deterministic effects in quantum mechanics
to someone who was not prepared to move beyond a 19th century mindset?
You have been given clear references to the sequence point rules, which
make it clear that there must be a sequence point between updates.
There is and can be no such sequence point when a location is updated
by passive one-sided communication and is later used in the process
that owns the data. Muttering about fences is irrelevant, because the
same applies even when you have them. The other MPI facilities were
carefully designed to ensure that there IS such a sequence point.
You are clearly thinking entirely in terms of trivial problems, where
failure is reliable and deterministic. Well, I have news for you.
Shared memory parallelism is not like that, and we didn't put all that
effort into the wording of the various standards for no reason (and I
have been involved in all of C, C++ and Fortran in this area). You
won't believe me, but you will see similar remarks here:
As I said, I have exposed those issues on many systems, including
the SGI Origin, SunFire F15K and others - and they were not bugs in
their implementations. Could I guarantee to repeat them on other
systems? Probably, but I would need to be hands-on to a suitably
large system and experiment with hammering it to hell and back until
I found a partially reproducible one. Even with access, I have much
better things to do. Can I give you a simple, repeatable one? Don't
be silly. The issue you have missed is that this happens with low
probability in real programs, but most people never track it down.
Enough is enough.
More information about the mpi-forum