Is there an established approach for getting fast compile times for WRF? For example if I'm doing code development, I might be more concerned with fast-compile times than fast model execution. In general I check out a dedicated node on a supercomputer, and try to compile in parallel with maybe '-j 12'. Often though the compilation seems to bottleneck to significant cpu consumption on a single processor. The model can take 1-2 hours for a full compile. This is a really difficult turn-around time, for development work. I've worked with numerous other gfd models and none take anywhere near as long. I'm guessing this has something to do with the extent to which WRF uses fortran modules.
slow compilation/build of WRF
- Thread starter sdurski
- Start date
You can try to use the new
github.com
compile_new and configure_new scripts to do compilation. This method generally alleviates some of the serialized compilation and can have a drastic improvement in compile times. This comment has a small breakdown of timing for GNU compilation for old vs new compilation methods:Breakout deallocation calls into simpler smaller files by islas · Pull Request #2070 · wrf-model/WRF
TYPE: enhancement KEYWORDS: intel, compilation, llvm, memory SOURCE: internal DESCRIPTION OF CHANGES: Problem: The Intel oneAPI compilers (and others like nvhpc) struggle with some of the larger (1...
github.com