Molcas Forum

yasuharushiraishi

Member

Registered: 2018-05-14

Posts: 6

CASSCF parallel calculation perfomance (calculation time, OrdInt file)

Hello,
I use latest OpenMolcas software (openmpi v.1.6.5, global array v.5.6.1, intel mkl 2017, gcc v.7.1.0 ),
and I want to do naphthalene CASSCF(10,10) calculation by using paralell computation.
But my CASSCF calculation by 1core is faster than calculation by 5core ( 1 core:24 s, 5 core:150 s),
and I think that this reason is caused by writing a large OrdInt file.

I Use「Direct」command, and do naphthalene HF calculation, calculation by 5core is 2 times
faster than calclation by 1 core. I think this reason is that large OrdInt file is not written
during the calculation.

Is there such command when I do CASSCF calculation?

my CASSCF calculation input file is as follows.
Best regards.

-input file of naphthalene-

&GATEWAY
coord
18

 C 0.00000000  -0.70895901   0.00000000
 C 0.00000000  -1.40219252   1.24866421
 C 0.00000000  -0.71153065   2.43773743
 H 0.00000000  -2.48463723   1.24629358
 H 0.00000000  -1.24697991   3.37782122
 C 0.00000000   0.70895901   0.00000000
 C 0.00000000   1.40219252   1.24866421
 C 0.00000000  -1.40219252  -1.24866421
 C 0.00000000   1.40219252  -1.24866421
 C 0.00000000   0.71153065   2.43773743
 C 0.00000000  -0.71153065  -2.43773743
 C 0.00000000   0.71153065  -2.43773743
 H 0.00000000   2.48463723   1.24629358
 H 0.00000000  -2.48463723  -1.24629358
 H 0.00000000   2.48463723  -1.24629358
 H 0.00000000   1.24697991   3.37782122
 H 0.00000000  -1.24697991  -3.37782122
 H 0.00000000   1.24697991  -3.37782122
group=XY XZ XYZ
basis=cc-pVDZ
&SEWARD
&SCF
&RASSCF
Title
Naphthalene cc-pVDZ
symmetry = 1
Nactel = 10 0 0
Spin = 1
Inactive = 9 0 0 6 0 7 7 0
Ras2 = 0 2 3 0 2 0 0 3

Ignacio

Administrator

From: Uppsala

Registered: 2015-11-03

Posts: 1,085

Re: CASSCF parallel calculation perfomance (calculation time, OrdInt file)

If all cores are in a single physical machine, sharing memory and disk access, it's very likely that you won't gain much with parallelization. Best parallel performance with OpenMolcas is typically obtained when each process has independent access to memory and disk.

Something you should definitely try is the RICD option in Gateway (but not that the absolute energies are not comparable with those obtained without it).

yasuharushiraishi

Member

Registered: 2018-05-14

Posts: 6

Re: CASSCF parallel calculation perfomance (calculation time, OrdInt file)

Dear Ignacio,
Thank you for your advice!

I try naphthalene CASSCF(10,10) calculation ,
and other molecule(50 atoms, symmetry CI) CASSCF(14,16) calculation
by using paralell computation (4 node×1 core, 8 node×1 core).

Naphthalene calculation speed is not up by 4 node and 8 node calculaion,
because I think naphthalene is small molecule.
But other molecule(50 atoms, symmetry CI) CASSCF(14,16) calculation
by 4 node is 2.5 times faster than calclation by 1 node,
and calculation by 8 node is 3.3 times faster than calclation by 1 node!

Naphthalene CASSCF(10,10) calculation time
              1node    4node     8node
seward    11s      14s       13s
scf            2s       6s        6s   
rasscf        4s       8s        8s
total       17s      28s        27s

Other Molecule(50 atom, symmetry CI) CASSCF(14,16) calculation time
               1node    4node     8node
seward      1.0h     0.5h       0.4h
scf            0.8h     0.3h       0.3h   
rasscf         13h      5.1h       3.8h
total        14.8h     5.9h      4.5h

Do you know my calculation performance is reasonable?
Can I speed up my calculation more?
Do you know CASSCF parallel calculation performance benchmark
of Molcas or OpenMolcas?

Best Regards.

Ignacio

Administrator

From: Uppsala

Registered: 2015-11-03

Posts: 1,085

Re: CASSCF parallel calculation perfomance (calculation time, OrdInt file)

An easy way to speed up most calculations is using RICD in GATEWAY, but note that you should not compare absolute energies with and without RICD.

yasuharushiraishi

Member

Registered: 2018-05-14

Posts: 6

Re: CASSCF parallel calculation perfomance (calculation time, OrdInt file)

Dear Ignacio,
Thank you for your advice!

I used RICD in GATEWAY, calculation is 2 times faster than direct method !
Energy difference is only 0.5 kcal/mol,
and scratch file size is 10 times smaller than direct medhod.

     1node(1core)   4node(4core)   8node(8core
Direct   14.8 h      6.0 h        4.4 h
RICD         7.9 h       3.2 h        2.3 h

I think RICD is very useful method.
Thank you very much!