Last-digit reproducibility, CCSD(T) optimization
allan.east at uregina.ca
Wed Mar 5 16:27:30 GMT 2003
Dear MOLPRO users;
We are running the downloaded MOLPRO2002.3 executable on a multi-CPU machine
with Linux (gnu, 2.4.18) and Athlon chips. We ran a test run,
CCSD(T)/cc-pVTZ optimization of C4H11+, on two different nodes, one asking
for 29Mw memory, and the other asking for 80Mw memory. Since we're using
the same executable, I was expecting the exact same output data, and the
first geometry, gradient, and Hessian do match exactly, but for the second
geometry, 3 of the dihedral angles differ by 1 in the last digit.
Ultimately, the optimizations converged but the final dihedrals differed by
10-4 or 10-5 degrees.
Yes this difference is small, and smaller than the noise we see when
comparing to results with our recompiled version of MOLPRO2002.3 with a
different BLAS library. However, in algorithms without Monte-Carlo-type
random number generators, an executable must give exactly reproducible
The only algorithmic difference I could see was that the 29Mw run had to do
multiple passes over integrals[?], while the 80Mw run did not. Does anyone
know if this difference could cause differences in computed values, and why?
The input is listed below.
University of Regina
More information about the Molpro-user