The DCLNO method [51] is a variant of the DC method. The computational efficiency is improved by introducing localized natural orbitals (LNOs) to span the subspace of each atom. The dimension of the resultant subspace is smaller than that in the DC method, leading to the reduction of computational cost. The LNOs are noniteratively calculated by a lowrank approximation via a local eigendecomposition of a projection operator for the occupied space. As shown in Fig. 19(a), introducing LNOs to represent the long range region of a truncated cluster reduces the computational cost of the DC method while keeping computational accuracy. The method can be applied to not only gapped systems, but also metallic systems as long as the size of truncated clusters is large enough, and typically clusters including more than 200 atoms might be chosen. The functionality is compatible with not only the collinear calculation, but also the noncollinear calculations.

As a first step for O() calculations of the DCLNO method, one can perform an O() calculation for Si crystal using an input file store in the directory 'work' as
% mpirun np 112 ./openmx Si8LNO.dat  tee si8lno.stdThe calculation was performed in 66 seconds using 112 cores on a Xeon cluster machine of 2.6 GHz. In order for you to start an trial calculation by the DCLNO method, an input file 'Si8LNO.dat' is available in the directory 'work'. Since as shown in Fig. 19(b) the three level parallelization has been implemented: atom level, spin level, and diagonalization level, it is expected that the method scales up to, e.g., 40000 CPU cores for a 1000 atom system in the parallel calculations, where we assumed 1000 atoms 2 (spinpolarized calculation) 20 CPU cores per node, resulting in that the product becomes 40000. The benchmark calculation of the multilevel parallelization will be shown later on. When you try to perform the hybrid parallelization, the following keyword has to be switched on:
scf.dclno.threading on # offonIn the hybrid parallelization the diagonalization at the bottom level will be parallelized by OpenMP. The computational accuracy and efficiency of the method can be controlled by the following keywords:
orderN.HoppingRanges 7.0 # 7.0 (Ang.) orderN.LNO.Buffer 0.2 # default = 0.2 orderN.LNO.Occ.Cutoff 0.1 # default = 0.1The role of the keyword 'orderN.HoppingRanges' is exactly the same as that in the DC method. For each atom a truncated cluster is constructed by pickingup atoms within a sphere whose radius is specified by the keyword 'orderN.HoppingRanges'. Though the proper choice of the parameter depends on systems, a serie of benchmark calculations implies that the accuracy is enough for not only gapped systems, but also metallic systems if 'orderN.HoppingRanges' is set so that the resultant truncated cluste can include 300 atoms. The setting might be regarded as a conservative choice to unsure the accuracy rather than efficiency. So, a compromising choice with respect to both accuracy and efficiency may be in between 200 and 300 atoms. The region in the truncated cluster where the PAOs are replaced by LNOs is determined by the keyword 'orderN.LNO.Buffer'. The 'orderN.LNO.Buffer=0.0' means that PAOs allocated on all the SNAN atoms are replaced by LNOs, while the PAOs on all the SNAN atoms remain unchanged in the case with the 'orderN.LNO.Buffer=1.0' which is equivalent to the DC method. As for the SNAN, please refer the subsection 23.4 'User definition of FNAN+SNAN'.

<Definition.of.Atomic.Species Si Si7.0s2p2d1 Si_PBE19 H H6.0s2p1 H_PBE19 Definition.of.Atomic.Species>the number of LNOs can be specified by
<LNOs.Num Si 4 H 1 LNOs.Num>In this case, the numbers of the LNOs are fixed to 4 and 1 for Si and H, respectively. To avoid a sudden change of the number of LNOs during geometry optimization and molecular dynamics simulations, it might be better to use 'LNOs.Num' rather than orderN.LNO.Occ.Cutoff. The comparison between the DC and DCLNO methods is shown in Fig. 20. Although the PAOs in the long range region are replaced by the LNOs, it is found that the accuracy is comparable to the DC method both in gapped and metallic systems. As an illustration for applications of the DCLNO method, we show in Fig. 21 radial distribution functions (RDF) of liquids for silicon, aluminum, lithium, and SiO. It turns out that in all the cases the DCLNO method reproduces well the results by the conventional O() diagonalization method, and that the obtained RDFs are well compared to other computational results [52,53,54,55].

In Fig. 22 the speedup ratio in the MPI parallelization of the DCLNO method is shown for nonspin
polarized calculations of a diamond supercell containing 64 atoms. Since the multiplicity of spin index is 1,
we see a nearly ideal behavior up to 64 MPI processes. Beyond 64 MPI processes the parallelization in the diagonalization
level is taken into account on top of the parallelization in the atom level.
