Details
Current setup
- ABACUS:
develop branch, CPU build
- Hardware: 2 nodes, 64 physical cores per node
- Parallel layout: 64 MPI processes with 2 OpenMP threads per process
- System size: 322 atoms
- Basis: LCAO DZP
- Exchange-correlation: PBE
- van der Waals correction: D3(BJ)
- k-points: Gamma-only
- Plane-wave cutoff:
ecutwfc = 100 Ry
- MD ensemble: NVT with a Nose–Hoover chain
- Temperature: 300 K
- Time step: currently 0.2 fs
- SCF threshold:
1e-6
- SCF solver: Broyden mixing
mixing_beta = 0.6
mixing_ndim = 8
- Charge density restarted from a previous converged calculation
Performance
With this setup, one MD step takes approximately 339 seconds. A typical step requires about 23 electronic iterations, with each iteration taking approximately 7 seconds.
I also added an ion-only restraint force to the code. Direct timing shows that the restraint calculation takes approximately 22 microseconds per MD step, so it has no measurable
impact on the total runtime and is not the performance bottleneck.
Task list for Issue attackers (only for developers)
Details
Current setup
developbranch, CPU buildecutwfc = 100 Ry1e-6mixing_beta = 0.6mixing_ndim = 8Performance
With this setup, one MD step takes approximately 339 seconds. A typical step requires about 23 electronic iterations, with each iteration taking approximately 7 seconds.
I also added an ion-only restraint force to the code. Direct timing shows that the restraint calculation takes approximately 22 microseconds per MD step, so it has no measurable
impact on the total runtime and is not the performance bottleneck.
Have you read FAQ on the online manual http://abacus.deepmodeling.com/en/latest/community/faq.html
Task list for Issue attackers (only for developers)