[Pw_forum] Nondeterministic rounding errors?

[Norbert Nemec]

I can actually answer my own question by the statement that I found on the web:

"MPI standard encourages but does not enforce determinism in reduction operations"

This makes clear that identical runs of PWSCF may well produce different results. Probably worth to keep this in mind for the future...


-------- Original-Nachricht --------
> Datum: Thu, 12 Aug 2010 13:46:46 +0200
> Von: "Norbert Nemec" <Norbert.Nemec.List at gmx.de>
> An: PWSCF Forum <pw_forum at pwscf.org>
> Betreff: [Pw_forum] Nondeterministic rounding errors?

> Has anyone seen something like this before:
> 
> I start two MD runs that are identical except for the initial velocities.
> Both are started on the same Cray machine on different but identical sets
> of cores, using the same binary. In my understanding, the first SCF loop
> should produce binary identical results. Yet - the wave function produced by
> pw2casino shows differences in the last bit of many floating point values.
> 
> Is there any point where non-determinism might affect rounding? Perhaps
> somewhere in the MPI communications?

-- 
GMX DSL SOMMER-SPECIAL: Surf & Phone Flat 16.000 für nur 19,99 ¿/mtl.!*
http://portal.gmx.net/de/go/dsl