In the Siesta 4.1 manual there is a clear description (Sec. 2.3.2) on how to utilize hybrid MPI+OpenMP using OpenMPI.
For your architecture/cluster it may be different, I would advice you contact your local HPC admin in that case. Den tir. 5. feb. 2019 kl. 22.03 skrev 郑仁慧 <z...@ustc.edu>: > Dear Nick: > Thank you very much for your reply. I can not understand your > explanation well. Would you like to give me an example using openmp or mpi? > Thank you again for your kind help. > Sincerely > Ren-hui Zheng > > > > > > > From: Nick Papior <nickpap...@gmail.com> > Date: 2019-02-04 15:07:10 > To: siesta-l <siesta-l@uam.es> > Subject: Re: [SIESTA-L] openmp > > Dear Ren-hui Zheng, > > Siesta can benefit from MPI, OpenMP or both. > The benefit may be briefly summarized as follows: > MPI has distribution across orbitals, hence a basic advice is that you get > performance scaling with # of cores == # of atoms (rule of thumb) > OpenMP also distributes across orbitals, but on a finer level. Here you > *may* get better scaling with more threads than atoms, however, generally I > would still suspect the best performance scaling up to # of cores == # of > atoms (rule of thumb). > > For the hybrid (MPI + OpenMP) the same thing applies, # of MPI-procs TIMES > # of threads per MPI-proc == # of atoms should give you reasonable scaling. > > Again, these are rule-of-thumbs as it also depends on your basis etc. > Note that OpenMP may be difficult to get correct scaling since affinity > plays a big role. How you place threads depends on your architecture and I > would advice you to do some benchmarks on your architecture to figure out > how to do best OpenMP. > > Den søn. 3. feb. 2019 kl. 22.01 skrev 郑仁慧 <z...@ustc.edu>: > >> Dear all: >> Can the siesta improves calculation speed when it optimizes >> molecular structure using openmp or mpi. When I optimize molecular >> structure with openmp or mpi, the calculation speed cannot be improved. I >> don not know the reason. Would someone like to explain it? Thank you very >> much for your help in advance. >> >> Sincerely >> Ren-hui Zheng >> >> > > -- > Kind regards Nick > > > -- Kind regards Nick
