Conventional scheme

Using the conventional diagonalization method, OpenMX Ver. 3.8 is capable of performing geometry optimization for systems consisting of 1000 atoms if several hundreds processor cores are available. To demonstrate the capability, one can perform 'runtestL2' as follows:

     % mpirun -np 128 openmx -runtestL2 -nt 4
  

Then, OpenMX will run with 7 test files, and compare calculated results with the reference results which are stored in 'work/large2_example'. The following is a result of 'runtestL2' performed using 264 MPI processes and 2 OpenMP threads on CRAY-XC30.

1	large2_example/C1000.dat	Elapsed time(s)= 767.64	diff Utot= 0.000000000560	diff Force= 0.000000006188
2	large2_example/Fe1000.dat	Elapsed time(s)= 8427.51	diff Utot= 0.000000006708	diff Force= 0.000000002148
3	large2_example/GRA1024.dat	Elapsed time(s)= 1418.26	diff Utot= 0.000000006759	diff Force= 0.000000002854
4	large2_example/Ih-Ice1200.dat	Elapsed time(s)= 304.88	diff Utot= 0.000000000209	diff Force= 0.000000000213
5	large2_example/Pt500.dat	Elapsed time(s)= 2552.29	diff Utot= 0.000000005013	diff Force= 0.000000000203
6	large2_example/R-TiO2-1050.dat	Elapsed time(s)= 1716.23	diff Utot= 0.000000000600	diff Force= 0.000000000200
7	large2_example/Si1000.dat	Elapsed time(s)= 814.97	diff Utot= 0.000000001037	diff Force= 0.000000000478

Total elapsed time (s) 16001.80

The quality of all the calculations is at a level of production run where double valence plus a single polarization functions are allocated to each atom as basis functions. Except for 'Pt500.dat', all the systems include more than 1000 atoms, where the last number of the file name implies the number of atoms for each system, and the elapsed time implies that geometry optimization for systems consisting of 1000 atoms is possible if several hundreds processor cores are available. The input files used for the calculations and the output files are found in the directory 'work/large2_example'. The following information is compiled from the output files.

No.	Input file	SCF steps	Elapsed time(s/SCF/spin)	Dimension
1	large2_example/C1000.dat	46	16	13000
2	large2_example/Fe1000.dat	343	16	13000
3	large2_example/GRA1024.dat	67	15	13312
4	large2_example/Ih-Ice1200.dat	32	6	9200
5	large2_example/Pt500.dat	190	16	12500
6	large2_example/R-TiO2-1050.dat	65	31	15750
7	large2_example/Si1000.dat	46	16	13000

The dimension of the Kohn-Sham Hamiltonian is of the order of 10000, and the elapsed time per SCF step is around 16 seconds for all the systems, implying that the difference in the total elapsed time mainly comes from the difference in the SCF iterations to achieve the SCF convergence of 10e-10 (Hartree) for the band energy.