[Pw_forum] Comparison of 3.1.1 and 3.2 (cvs)

[Paolo Giannozzi]

On Nov 23, 2006, at 9:46 , Huiqun Zhou wrote:

> There are two problems here:
> (1) The differences between wall time and CPU time with version
>  3.1.1 is small and seems reasonable while that between the two
> with cvs version is relatively large. Why?

maybe because they are not the same quantity in the two
cases. There have been some changes to the CPU and
wall time counters in the cvs version with respect to the
previous one. Note that printed cpu and wall times may
not always be what they are intended to be, depending
on the compiler. On at least some versions of the Intel
compiler, a call to "standard" routine cpu_time returns
the wall time, for instance.

> (2) Version 3.1.1 is much faster than the cvs version as the former
>  took only 13 scf iterations to reach convergence while the latter
>  needed 16 iterations.

the number of scf iteration may occasionally be sensitive to
minor numerical changes. There hasn't been any major change
to the self-consistency algorithm, so I don't expect any major
or generalized change in the overall performances.

> Of course, it's just a cvs version and development is in progress.
> I hope that developers have already noticed the difference.

personally I haven't noticed anything: benchmarking is not my
preferred activity (I would even say, like a famous physicist
about chemistry: "nothing I am more allergic to". CPU and wall
times are irreproducible on modern machines: I have several
times noticed variations of 10 to 15% with no apparent reason).

> Thanks to all developers for their great effort!

thank to you for your report. I hope that somebody will
have a look at it and use it to improve performances.
Note however that unless a clear reason (if any) is found
for the performance degradation you observed, it will be
hard to do any better

Paolo
---
Paolo Giannozzi, Democritos and University of Udine, Italy