Alessandro Delo
New member
Hi all
We are running several experiment by using WRF (v. 4.4.1) coupled with WRF_HYDRO (v. 5.2.0). After several attempts, we understood that it was not possible to run WRF+WRF_HYDRO in hybrid mode (i.e. OpenMP works for WRF but not with WRF_HYDRO). We ran a series of tests to try to improve efficiency (i.e. time calculation), and we achieved the following:
1. We tried tiling, but it doesn't seem to work for WRF_HYDRO;
2. Compilation options: we compiled with mpiifort (Intel compiler) but saw no improvements with respect the attempt we did by compiling with mpif90;
3. Quilty option to reduce output file writing time, but this also doesn't work with unpaired WRF;
4. By subdividing N tasks, with NtaskX << NtaskY, we obtained a reduction in CPU time;
5. When submitting a job, not using all the cores of a node, but only a part of them (32 instead of 36) leads to a sensible reduction of time;
All these tests reported here were compiled by activating the NetCDF4 option, which allowed us to reduce the output size but, at the same time, increased writing times.
In conclusion, for the domain we are analyzing, we managed to reduce the calculation times from approximately 9 min per day to 6 min per day. The goal would be to halve these times in view of a coupled climate simulation of around 100 years.
We attach the configure.wrf we are using, the Job submission script and some info on the supercomputer we are using.
If you have any ideas to suggest on how to reduce calculation times, this would be of great help to us.
Thanks in advance and sorry for the long message!
Alessandro
We are running several experiment by using WRF (v. 4.4.1) coupled with WRF_HYDRO (v. 5.2.0). After several attempts, we understood that it was not possible to run WRF+WRF_HYDRO in hybrid mode (i.e. OpenMP works for WRF but not with WRF_HYDRO). We ran a series of tests to try to improve efficiency (i.e. time calculation), and we achieved the following:
1. We tried tiling, but it doesn't seem to work for WRF_HYDRO;
2. Compilation options: we compiled with mpiifort (Intel compiler) but saw no improvements with respect the attempt we did by compiling with mpif90;
3. Quilty option to reduce output file writing time, but this also doesn't work with unpaired WRF;
4. By subdividing N tasks, with NtaskX << NtaskY, we obtained a reduction in CPU time;
5. When submitting a job, not using all the cores of a node, but only a part of them (32 instead of 36) leads to a sensible reduction of time;
All these tests reported here were compiled by activating the NetCDF4 option, which allowed us to reduce the output size but, at the same time, increased writing times.
In conclusion, for the domain we are analyzing, we managed to reduce the calculation times from approximately 9 min per day to 6 min per day. The goal would be to halve these times in view of a coupled climate simulation of around 100 years.
We attach the configure.wrf we are using, the Job submission script and some info on the supercomputer we are using.
If you have any ideas to suggest on how to reduce calculation times, this would be of great help to us.
Thanks in advance and sorry for the long message!
Alessandro
Attachments
Last edited: