I have successfully compiled both with mpif90 (Tried GNU MPICH2, MPICH2 1.4, Intel OpenMPI) and verified that the correct libraries were in the path for each.
When I run VASP compiled this way(just by calling vasp not mpirun) and checked the used resources using top it claims to be using 2000%+ cpu (48 core AMD opteron chipsets). The run completes without error. But as I submit it to multiple processors using mpirun it slows because all threads are looking for the same about of cpu %(which corresponds to multiple cores).
The results are consistent with a working version on another machine.
I considered that it could be a problem with mpirun but I have other software compiled (LAMMPS, GPAW, QE) all of which function correctly.
I am baffled as to why this is happening. Perhaps I am missing an environmental variable?
Parallelization issue
Moderators: Global Moderator, Moderator
-
westover
Parallelization issue
Last edited by westover on Tue Dec 27, 2011 5:46 am, edited 1 time in total.
-
westover
Parallelization issue
I have recompiled with a generic makefile provided with ifort as opposed to mpif90 and I am still seeing the same issue.
Last edited by westover on Wed Dec 28, 2011 10:06 am, edited 1 time in total.
-
westover
Parallelization issue
It appears to be a library issue
Last edited by westover on Wed Dec 28, 2011 11:21 am, edited 1 time in total.
-
westover
Parallelization issue
Issue resolved. For all of those who have similar issues in the future it should be noted that the new gotoblas2 libraries compile as multithreaded by default and it should be compiled with make USE_THREADS=0
Additionally some time should be taken by the admin to look at the makefiles for the intel compiler to touch up the libraries for the more recent versions of IFORT
Additionally some time should be taken by the admin to look at the makefiles for the intel compiler to touch up the libraries for the more recent versions of IFORT
Last edited by westover on Wed Dec 28, 2011 11:45 am, edited 1 time in total.