Parallelization

The code spends most of the time in LAPACK diagonalization routines (dsyev, dsyevr, zheev, etc.) and in level-3 BLAS matrix-matrix multiplication routines (dgemm, zgemm). The execution time can be significantly reduced by using a high-quality multithreaded BLAS/LAPACK libraries, such as Intel MKL. Make sure that the relevant environment variables (e.g. OMP_NUM_THREADS, MKL_NUM_THREADS, MKL_DYNAMIC) are set appropriately. Good choices are e.g. 4 or 8 threads.

The calculations over the different values of the twist parameter \(z\) (different interleaved discretization grids) are done in parallel using MPI parallelization. Good choice for the number of MPI processes is the value of Nz, i.e., the number of different grids.

Finally, the diagonalisation of matrices can be OpenMP parallelized. This is, however, only beneficial for large-scale multi-orbital calculations and is quite seldom used. This is configured using the parameter diagth in nrg_params_t (“low-level NRG parameters”), the default value being 1 (diagonalisations performed in series).