[Pw_forum] timing vs QE version vs pools

Paolo Giannozzi giannozz at democritos.it
Tue Mar 10 18:25:09 CET 2009

Giovanni Cantele wrote:

> Can it be due to wrong settings of my cluster?

there is no evidence that anything is wrong in your
clusters (or more wrong than in the average cluster).

I don't see any reason why the new version should
be slower than the old one. I never noticed what
you report. Not that I trust benchmarks. Years ago,
on a IBM 320h machine, I found a case in which the
addition of a single unused variable  reproducibly
increased the cpu time from 27' to 30'. I have also
seen a case in which the same executable, on the same
empty machine, had random 10-15% variations in the
execution time for no apparent reason.

About parallel distributed diagonalization: the optimal
number of processors used to be selected at run time,
but it was slow and not always accurate. Since it
was found that in most cases parallel diagonalization
was faster or not worse than the serial one, it was
chosen as default. Of course it is impossible to choose
defaults that are optimal for everybody.

Paolo Giannozzi, Democritos and University of Udine, Italy

More information about the users mailing list