[CP2K:2214] Re: determinism of CP2K runs

Noam Bernstein noam.be... at gmail.com
Mon Aug 17 19:39:52 CEST 2009


OK - here's the final word on my determinism issue.

Apparently with the MPI/ScaLAPACK version I have now
(and unlike some versions I was using earlier, or perhaps
because CP2K uses different routines), the BLACS reduction
routines are not, by default, guaranteed to be repeatable.
In particular, pdsyevd, which computes eigenvalues, gives non
repeatable results.  Using
   call BLACS_SET(DESCZ(CTXT_), 15, 1)
at the beginning of pdsyevd fixes this.  I'll probably add a
patch to CP2K to make guarantee repeatability for all BLACS
contexts.  If anyone else wants it, I'll be happy to contribute it.

                Noam



More information about the CP2K-user mailing list