large systems

[willphy123]

Dear Dr. Tadano,
I have a large system that has ~350 atoms. I created a 221 supercell to produce displacement patterns. Because the system has no symmetry, there are a lot of displacements. I want to ask if alm can do parallelization to create displacement patterns? The calculations are very slow. Can you give me advice to accelerate creating them?
Thanks
Will

[ttadano]

Do you want to parallielize the calculations of atomic forces for the 221 structures? If so, that can be done by submitting several jobs to your supercomputer nodes, each of which works on a different subset of the structures.

[willphy123]

Thanks for your quick reply. No, I'm not calculating forces for 221 structures. Instead, I created a 2 by 2 by 1 supercell, not 221 structures. Actually, I am at the very beginning of obtaining displacement patterns. The process of creating displacement patterns is very slow. It is still not finished after two hours.
So I want to ask if alm can do parallelization to obtain displacement patterns.

[ttadano]

ALM is partly parallelized using OpenMP. If you are using a multi- or many-core CPU and the binary is compiled with -fopenmp or -qopenmp option, the OpenMP parallelization is activated by default. Please see the information in the header part of the log file.

If OpenMP parallelization is already active, you may need to reduce the cutoff radius or use a faster CPU. Since your primitive cell is already fairly large, I think you don't need to use a 2x2x1 supercell; the primitive cell calculation would be sufficient, I guess.

[willphy123]

OK. I will try the primitive cell.

Output:

---

+-----------------------------------------------------------------+

•                     Program ALM                             +
  

•                         Ver.  1.4.1                         +
  

+-----------------------------------------------------------------+

•                     Program ALM                             +
  

•                         Ver.  1.4.1                         +
  

+-----------------------------------------------------------------+

+-----------------------------------------------------------------+

•                     Program ALM                             +
  

•                         Ver.  1.4.1                         +
  

+-----------------------------------------------------------------+

+-----------------------------------------------------------------+

•                     Program ALM                             +
  

•                         Ver.  1.4.1                         +
  

+-----------------------------------------------------------------+

Number of OpenMP threads = 28

Number of OpenMP threads = 28

Job started at Thu Jun 30 05:47:33 2022

Job started at Thu Jun 30 05:47:33 2022

Number of OpenMP threads = 28

Input variables:

General:
PREFIX = c_221
MODE = suggest
NAT = 1456; NKD = 1
PRINTSYM = 0; TOLERANCE = 0.001
KD = C
FCSYM_BASIS = Lattice
PERIODIC = 1 1 1
MAGMOM =
FCS_ALAMODE = 1; NMAXSAVE = 5
FC3_SHENGBTE = 0
FC2_QEFC = 0
HESSIAN = 0
FC_ZERO_THR = 0

Interaction:
NORDER = 1
NBODY = 2

DBASIS = CART

---

Input variables:

General:
PREFIX = c_221
MODE = suggest
NAT = 1456; NKD = 1
PRINTSYM = 0; TOLERANCE = 0.001
KD = C
FCSYM_BASIS = Lattice
PERIODIC = 1 1 1
MAGMOM =
FCS_ALAMODE = 1; NMAXSAVE = 5
FC3_SHENGBTE = 0
FC2_QEFC = 0
HESSIAN = 0
FC_ZERO_THR = 0

Interaction:
NORDER = 1
NBODY = 2

DBASIS = CART

[LePingKYXK]

Hi Will,
May I ask whether the parallelization calculation could be run if I install the libraries via conda? Or is the parallelization performed only when libraries are installed from the source?

[LePingKYXK]

ALM is partly parallelized using OpenMP. If you are using a multi- or many-core CPU and the binary is compiled with -fopenmp or -qopenmp option, the OpenMP parallelization is activated by default. Please see the information in the header part of the log file.

If OpenMP parallelization is already active, you may need to reduce the cutoff radius or use a faster CPU. Since your primitive cell is already fairly large, I think you don't need to use a 2x2x1 supercell; the primitive cell calculation would be sufficient, I guess.

Dear Dr. Tadano,
As you mentioned that the compile options -fopenmp or -qopenmp, is that mean the parallelization calculation could be run only if I compile the libraries (the Boost C++, Eigen3, and spglib libraries) from the source and compile alamode through makefile?

Can I run parallel calculations if I install the libraries via conda and then compile alamode via makefile?

By the way, the link of spglib has changed to https://github.com/spglib/spglib. The link (https://alamode.readthedocs.io/en/latest/install.html#step-1-install-all-required-libraries) needs to be updated.

[ttadano]

The OpenMP parallelization will also be enabled by the conda build if the compiler supports OpenMP. Please check if the CMake output contains the following line:

-- Found OpenMP: TRUE (found version "4.5")

If the flag is TRUE, OpenMP is enabled.