Parallel version slower than the serial one
Posted: February 14th, 2020, 5:54 am
Hi,
I am facing this issue with the parallel version.
I have compiled it with mpich by amending the Makefile.inc file, and it successfully generated the parallel executable (hycm_std). I am now running a test case both with the serial code (as usual) and parallel. The serial code took less than 8 seconds, the parallel one 58 seconds.
I have launched the parallel version with this command:
time sh ../../../../../hysplit.v4.2.0/exec/run_mpi.sh 10 hycm_std
after amending run_mpi.sh to point at the location of mpich executables and HYSPLIT executables (see attached). Note that the output files from each processor (MESSAGE, VMSDIST) have been generated, and I also cross-checked it was actually using 10 processors during runtime.
I have noticed that other people experienced the same issue; I also raised this point some time ago (with no answers so far), but I hoped to solve this with the installation of the latest version.
Is there anything else I need to do to make the parallel version really efficient?
Thanks
Regards
Fabio
I am facing this issue with the parallel version.
I have compiled it with mpich by amending the Makefile.inc file, and it successfully generated the parallel executable (hycm_std). I am now running a test case both with the serial code (as usual) and parallel. The serial code took less than 8 seconds, the parallel one 58 seconds.
I have launched the parallel version with this command:
time sh ../../../../../hysplit.v4.2.0/exec/run_mpi.sh 10 hycm_std
after amending run_mpi.sh to point at the location of mpich executables and HYSPLIT executables (see attached). Note that the output files from each processor (MESSAGE, VMSDIST) have been generated, and I also cross-checked it was actually using 10 processors during runtime.
I have noticed that other people experienced the same issue; I also raised this point some time ago (with no answers so far), but I hoped to solve this with the installation of the latest version.
Is there anything else I need to do to make the parallel version really efficient?
Thanks
Regards
Fabio