Mixed Shared/Distributed Memory Parallelization

The two parallelization schemes described above are implemented in such a way that they don't interfere and can be used independently or at the same time. This can be especially beneficial if the MPI parallelization is penalized by communication bandwidth limitations (e.g. when having many CPU cores per node, or when using a gigabit ethernet interconnect).

Due to the high efficiency of the MPI parallelization in CPMD for a small to medium number of nodes and the fact all modern MPI libraries are able to take advantage from shared memory communication, using the plain MPI parallelization is usually the fastest option. The mixed shared/distributed memory parallelization is of most use, if you run a job on a large number of SMP nodes, when the distributed memory parallelization has reached its scalability limit or the network is overloaded (see above). To learn more about the mixed parallelization scheme of CPMD consult reference [6].