Contents

12 Performance of sapt2012

The computational cost of sapt2012 calculations depends on the system size in the following way. The integral/SCF calculation time scales at most as (o + v)⁴, with a somewhat lower power expected for large systems. When the AO-based algorithm for the four-virtual diagram in CC is used (which is the default in sapt2012.2 when atmol1024 or molpro are used as the integral/SCF program), the transformation step scales as o(o + v)⁴ and the CCSD step as o²(o + v)⁴. Otherwise, the scalings of the transformation and CCSD steps are proportional to v(o + v)⁴ and o²v⁴, respectively. Finally, the scaling of the perturbation theory code strongly depends on the required level of theory: o³v² for SAPT(0), o²v³ for SAPT2, and o³v⁴ for the full SAPT. For a general orientation, Table 2 contains timings from an Opteron 252 Linux system, obtained with the sapt2012 suite compiled in the 64-bit mode with Portland pgf90 compiler. The examples chosen are described in Sec. 13 under the labels ArHF_DCBS, CO2D_MCBS, and C6H6_H2O_DCBS. For the sake of completeness, the amounts of memory and disk space required in each calculation are also reported (see Sec. 10.7 for a general discussion of these two issues). The memory requirements of the integral/SCF calculations are omitted because they are negligible compared to the other steps. Table 2 is restricted to regular SAPT calculations on a single core. See Ref. 57 for a detailed analysis of the SAPT(DFT) performance and Ref. 9 for a discussion of the parallel SAPT scaling.