WRF4DVAR fails with intel LLVM compilers #1957
Comments
|
@HathewayWill We are aware of the compilation issue for DA and Chem code using the newest Intel compiler. |
|
Wasn't sure. Didn't see it in the GitHub discussion. |
|
@HathewayWill I have added a note in the release note. |
|
I'll try to find a solution. I have a lot of free time. |
|
Good morning, I was sucessfully able to get Intel LLVM to install WRFPLUS/WRF4DVAR by adding the following commands: One thing I did notice for WRFPLUS and 4DVAR there is a big memroy leak somewhere during installation. I have 64GB of RAM and 64GB of SWAP RAM and it was maxing out my physical RAM and then half of my SWAP. I didn't get to see which module was causing it but I think I saw @islas mention there was a process that needed -j 1 somewhere in the documentation for the release notes. Hope this helps you. |
|
Here's the output during the memory leak. Also affects WRF chem |
|
@HathewayWill Can you make a PR with your fixes for WRFDA/WRFPlus compilation with Intel-OneAPI compiler? |
|
I don't know how to do a PR so I was letting NCAR staff look at my comments and files and let them do it. |
|
@liujake the memory leak is another problem I don't know how to fix but it's documented in the zip file. |
|
@HathewayWill I'm not sure I follow how adding these flags: It sounds like these flag additions were independent of the memory leak issue, is that correct? |
|
@islas I'm current sick with COVID and in isolation for 2 weeks. Let me get back to you when I feel better. |
Yes the memeory leak is different. I will rerun the installation without the flags added to show the errors that popped up |
|
So here are two log files from WRF v4.5.2 that doesn't compile correctly when those flags are not included. |
|
Thanks @HathewayWill It may be a little difficult to tell, but a sure way to identify it is to look at when the real compile command happens vs when the rm command happens with respect to other files (rm is the first command in the WRF makerule for these files):
PR #1950 aims to fix these issues - if you search the Log 2 is a little more confusing, I'm not too sure about what the issue is there and I can't definitively rule out an environment issue - however the flags in question should not affect whether MPI or not could be found. In the end, I am confident log 1 is not truly remedied by the flags and is most likely just adjusting the compilation race condition that exists (i.e. getting lucky) and log 2 shouldn't be affected by those flags. |
|
okay that's good information, so those flags that were turning errors into warnings was just a lucky guess then? |
|
Yes, though I haven't taken a look at the logs you posted in that issue the problem described matches the erroneous behavior pretty well. I suspect it may help, though #1950 only affects the WRF core objects, so if dependencies under the chem or da are missing those will still be issues. |
@islas so that will involve more detailed dives into the logs. Let me know which tests I can do to help because I think those log files for DA and Chem used the flags to make it work. I can always rerun it without them |
|
Here are the log files @islas without any flags added for WRFDA 4DVAR |
|
Are you seeing these issues on either the latest updates from develop (9e265af) or the current release candidate (release-v4.6.0)? These now include build dependency fixes and syntax/flag updates for the new Intel oneAPI compilers for WRF, WRFDA, and WRF-Chem |
|
Is there a .tar file for these? I'm not really familiar with how to pull with github. |
|
@HathewayWill Do this: |
Thank you @weiwangncar I'll try it today |
|
Good morning @weiwangncar @islas @kkeene44 @mgduda Here are the log files for each issue and their update. Tested on Ubuntu 22.04.4, 64GB of physical RAM 64GB of SWAP RAM, release candidate 4.6.0 #1992 #1981 #1967 #1957 The Memory leaks in chem and wrfplus maxed out my 128GB worth of RAM and shut down the computer. Happens at the same exact spot on each compilation which is confusing to me. |
|
I've taken a closer look at the issue, and as far as I can tell this is not a memory leak (though it does take an exorbitant amount of memory) on the WRF-side of things. I can't say for certain whether it is an Unfortunately, this is fundamental "feature" of the WRF autogenerated code from the registry and will require splitting some of the larger includes into separable files. I've already attempted this with Fortran submodules with limited success as not all WRF-supported compilers implement this well, thus a better approach for splitting the code up would need to be investigated. For reference, attached are two outputs : one from the make build and another from the cmake build - both of which show massive spikes in compilation of |
|
Anything I can do on my side to help this? |
|
You tried setting FCFLAGS and CFLAGS with no -03 correct? |
|
Correct, |
I have also tested different versions of the MPI compiler commands mpiifx all of them do the same thing with and without optimizations @islas |
Quick question. Why would having the warnings suppressed allow the WRF to install without any issue? |
|
reopening with new issue. @islas @weiwangncar @mgduda see attached log files for errors, to many errors to list. The files crashed the pc |
Describe the bug
Intel LLVM compilers for WRF da fails to build all the required exe files.
libufr fails to build.
To Reproduce
fails.zip
using option 40 for intel llvm dmpar
Expected behavior
expected 43 exe in /varr/da
expectted 1 exe in var/obsproc/src
got 42 exe in /var/da
got 0 iexe in /var/obsproc/src
Screenshots
If applicable, add screenshots to help explain your problem.
Attachments
works.zip
fix: add the following flags for llvm compilers to CFLAGS
CFLAGS_LOCAL = -w -O3 -ip -Wno-implicit-function-declaration -Wno-incompatible-function-pointer-types#-xHost -fp-model fast=2 -no-prec-div -no-prec-sqrt -ftz -no-multibyte-chars # -DRSL0_ONLYAdditional context
Add any other context about the problem here, such as:
The text was updated successfully, but these errors were encountered: