[CP2K:2306] Re: CP2K build with Intel Compiler and MKL

[Ondrej Marsalek]

Axel,

thanks for the comments.

On Fri, Sep 25, 2009 at 17:44, Axel Kohlmeyer <akoh... at gmail.com> wrote:
>
> one comment on this issue. for the example mentioned in the
> discussion,
> it looks to me, by comparing the output from different compilations,
> as
> if the regtest is done in a way that at some point there is a
> numerical
> instability in the code, and that for some families of compilers, the
> result flips one way (where the run finishes), and for others the
> other
> way (where it crashes). so i would argue that it is not always
> justified
> to blame the compiler outright, even if there are some compilers that
> get it "right".

This sounds quite scary to me. CP2K is notoriously tricky to compile
and so far for me the only reasonable way to tell whether I have a
working binary was to see if it can run the regtests successfully.
What you are saying means that this is not a very good method.

I do realize that I can never be sure that the build is 100% good.
However, in this case, I can't even tell if the build is wrong. It
crashes on the tests, but perhaps it is due to unstable tests,
perhapas it is due to miscompilation. This is really nasty.

Let me ask a slightly more general question. Does anyone routinely run
a build that does not pass the tests? Does it work well outside of the
tests?

> apart from the fact that the input may cause these
> instabilities
> due to the nature of the underlying algorithms, one also has to
> consider
> to consider that these kinds of instabilities can be reduced by
> changing
> the code to be more careful, and that for package software as large
> and
> as complex as cp2k, it is almost bound to happen in some areas of the
> code, due to the different ways how different people write code.
>
> that being said, i have also made the experience that particularly on
> itanium processors, the intel compilers are much more unstable than
> elsewhere. the good news is, that it looks as if the itanium is going
> the way of the dodo and that it soon will not be worth worrying about
> it.

I do all of this on the x86-64 platform. Also, it certainly is not
only the Intel compiler that has (or did have) trouble producing a
build that would pass the tests.

Ondrej