Slow runs with new WRF v4.6

[janac4]

Hello everyone,

I have compiled WRF 4.6.0 with the configure option dm+sm on an aarch64 architecture. After doing some tests, I found out that a domain with 400x300 cells with the new WRF version, needs a much bigger machine and more processors than the same domain ran with version 4.5.1.
Does it make sense?

Thanks for your help.

 

[haiqingsong]

WRF4.6 was compiled with same setup？

 

[William.Hatheway]

@janac4 did you change any of the compilers?

 

[janac4]

@William.Hatheway @haiqingsong
In version 4.5.1 we used gcc, g++ and fortran version 10, but in this new WRF 4.6 we used version 11 of the compilers.

 

[Ming Chen]

if you compile WRF in dmpar mode, can you find similar 'slow' run?
Usually the model compiled using intel runs faster than that compiled using GNU.

 

[janac4]

We have compiled the same WRF version with "dmpar" mode but found the same performance issues. We can't use Intel compilers as we are working with aarch64.
Does anyone have any other ideas?

 

[Ming Chen]

We don't have such issues ....
Have you talked to your computer manager ? I am suspicious this may be a machine issue.

 

[janac4]

I talked to my computer manager and we both think our machine is working properly. We tried switching from MPICH to OpenMPI and both give similar issues. When executing the run with 8 cores both fail with a segmentation fault. With MPICH we can increase that number of processors to 48, and run the simulation with a really bad performance. However with OpenMPI whe can't even execute the run without errors, it always throws a segemntation fault or a problem with the number of available slots (as shown in the image).

With a smaller domain (160x100 cells) the model runs without problem and we have noticed and increase of computation speed in comparison with WRF 4.5.1 (so we guess the compilation is correct). However, when increasing the model domain, it cracks every time.