Devs,
I wanted to bring to your attention a notable performance issue with the master branch of PFLOTRAN, which now builds with PETSc “main.” The compilation times are significantly slower compared to the previous version built with PETSc 3.21. Below are the results
observed on my Linux box:
GNU Fortran
|
PETSc main
|
PETSc 3.21
|
1 process
|
14:54
|
6:37
|
2 processes
|
12:15
|
4:04
|
4 processes
|
6:17
|
2:48
|
8 processes
|
5:15
|
2:21
|
16 processes
|
5:02
|
2:20
|
The issue seems to stem from the module compilation process on the PETSc side, which is now considerably slower. On the bright side, the Intel compilers build PFLOTRAN quickly, and it may be worth exploring this option further. However, please note that regression
testing will still need to be conducted using GNU compilers.
Regards,
Glenn