Describe the bug
An inconsistent behavior is found for mpirunning ABACUS with 2 processors among running serially, with 2, 4, and 8 cores.
The system tested is a carbon atom in a cubic box of 1.5-Angstrom edges. The calculation uses 20 Ryd as the kinetic energy cutoff, and 3 grid points along each dimension, int total 27, are used. Numerical orbitals are used.
The r2SCAN functional is used, with and without mixing kinetic-energy density, to obtain the electronic structure.
Two commints were tested, one is called new, 35448cb, commited on March 31st, 2025. The other commit is called old, 2f15c55, commited on November 12th, 2024.
For the new commit, when running with "-np 2", the SCF calculation does not converge, whereas serial, "-np 4" and "-np 8" are the same to each other to 10^-8 eV. The SCF loops show that the calculations agreeing with each other have their densities converged at the 14th SCF step, whereas the "-np 2" calculation does not manage to converge here. (see serial.log, np2.log, np4.log, and np8.log is the attached zipped folder)
I also attach the calculations 1 core and 2 cores with the old code. The same error is also found in the calculation (see log files in old_serial.log, and old_np2.log in the attachment).
However, if kinetic-energy density is not to be mixed in the calculation by commenting out mxing_tau, the old code give the same results for single- and two-core calculations.
test_218.zip
Expected behavior
the converged SCF energy should be the same regardless of how many cores are used in the calculation when mixing kinetic energy density.
To Reproduce
No response
Environment
No response
Additional Context
No response
Task list for Issue attackers (only for developers)
Describe the bug
An inconsistent behavior is found for mpirunning ABACUS with 2 processors among running serially, with 2, 4, and 8 cores.
The system tested is a carbon atom in a cubic box of 1.5-Angstrom edges. The calculation uses 20 Ryd as the kinetic energy cutoff, and 3 grid points along each dimension, int total 27, are used. Numerical orbitals are used.
The r2SCAN functional is used, with and without mixing kinetic-energy density, to obtain the electronic structure.
Two commints were tested, one is called new, 35448cb, commited on March 31st, 2025. The other commit is called old, 2f15c55, commited on November 12th, 2024.
For the new commit, when running with "-np 2", the SCF calculation does not converge, whereas serial, "-np 4" and "-np 8" are the same to each other to 10^-8 eV. The SCF loops show that the calculations agreeing with each other have their densities converged at the 14th SCF step, whereas the "-np 2" calculation does not manage to converge here. (see serial.log, np2.log, np4.log, and np8.log is the attached zipped folder)
I also attach the calculations 1 core and 2 cores with the old code. The same error is also found in the calculation (see log files in old_serial.log, and old_np2.log in the attachment).
However, if kinetic-energy density is not to be mixed in the calculation by commenting out mxing_tau, the old code give the same results for single- and two-core calculations.
test_218.zip
Expected behavior
the converged SCF energy should be the same regardless of how many cores are used in the calculation when mixing kinetic energy density.
To Reproduce
No response
Environment
No response
Additional Context
No response
Task list for Issue attackers (only for developers)