Modern Intel compilers struggle to compile McLachlan resulting in extremely long build times

[Gabriele Bozzola]

As mentioned in ET Call on 09/02/22.

I am compiling a subset of the latest release of the Einstein Toolkit with one of latest version available of Intel oneAPI (2021.5). I am attaching thornlist and optionfile.

I see warnings like this:

/home/sbozzolo/storage/codes/einstein_toolkit/Cactus/configs/sim/build/ML_BSSN/ML_BSSN_EvolutionInteriorSplitBy2.cc(49): warning #16219: Some OpenMP processing was skipped to constrain compile time. Consider overriding limits (-qoverride-limits).

which essentially informs me that the compiler cannot optimize the code within a time frame that is heuristically determined.

As a result, compilation takes a very long time (several tens of minutes). I have seen this on other supercomputers with compilers from 2019.

‌

(I don’t seem able to attach two files, so I am attaching only the option list. )

[Gabriele Bozzola]

  <div class="preview-container wiki-content"><!-- loaded via ajax --></div>
  <div class="mask"></div>
</div>

</div> </form>

[Steven R. Brandt]

Did you experiment with setting -qoverride-limits` ?

[Gabriele Bozzola]

My understanding is that it does the opposite of what I want: it lets the compiler try to optimize as much as it wants (https://www.intel.com/content/www/us/en/develop/documentation/cpp-compiler-developer-guide-and-reference/top/compiler-reference/compiler-options/compiler-option-details/advanced-optimization-options/qoverride-limits-qoverride-limits.html)

‌

[Erik Schnetter]

The usual remedy is to reduce the optimization level for that particular file. This is, of course, not really a good idea since that file contains one of the most expensive routines.