Counting the RelCC memory consumption¶
It is of great importance to know the exact number of the memory needed for a high memory and time demanding RelCC job.
The first step is the SCF run with saving molecular orbitals, preferably as DFPCMO text file. Afterwards, using obtained molecular orbitals, prepare the active space for MOLTRA and RelCC.
MOLTRA & RelCC¶
Equip the MOLTRA input with the .NO4IND keyword and RelCC input with the .COUNTMEM keyword. Run SCF part with previously saved converged molecular orbitals, with one iteration. The job passes quickly through MOLTRA integrals transformations as it produces only two-index quantities in the MRCONEE file.
In the RelCC module, the program counts the memory need and leaves.
Demonstration of this feature is in the DIRAC test/count_cc_memory, see test
We found the almost minimal setting - via the –ag parameter flag - for the memory counting run:
pam --noarch --gb=0.20 --ag=0.366 --inp=H2O.ae4z.x2c.scf_countmem-relcc.inp --mol=H2O.xyz --put "DFPCMO.H2O.x2c.ae4z=DFPCMO"
We are to look the “Peak memory usage” to get reasonable estimate of dynamic memory:
Predicted RelCC memory demand: 0.335 GB Peak memory usage (Gb) : 0.363
Next step is the sharp run based on the previous RelCC memory conuting. The serial job needs at least 0.366 GB, rounding to 0.4 GB of the total memory:
pam --noarch --gb=0.20 --ag=0.366 --inp=H2O.ae4z.x2c.scf_relcc.inp --mol=H2O.xyz --put "DFPCMO.H2O.x2c.ae4z=DFPCMO"
We also perform the parallel run, which consumes at least 16*0.366 = 5.856 GB of the memory, rounding to 5.9 GB. This amount of memory must be free at the running machine equiped with at least 16 CPUs:
pam --noarch --mpi=16 --gb=0.20 --ag=0.366 --inp=H2O.ae4z.x2c.scf_relcc.inp --mol=H2O.xyz --put "DFPCMO.H2O.x2c.ae4z=DFPCMO"
Have a look at the output file. We search there the peak memory usage which is cca 99.2% of the –ag parameter.
Predicted RelCC memory demand: 0.335 GB Peak memory usage: 0.363 GB
Therefore in practical RelCC memory demanding calculations watch for the Peak memory usage value and set the –ag parameter accordingly. This may take few iterations, especially when you want to find the real memory minimum. Don’t forget to multiply your memory amount with the number of threads in parallel calculations.