kpar	n_atoms	speedup	mpi_ranks	structure	best_wall_s	worker_cores
1	8	1	1	NiO	389.9	8
8	8	6.2	8	NiO	62.7	8
16	8	9.6	16	NiO	40.7	16
32	8	17.8	32	NiO	21.9	32
1	32	1	1	FeCo 2x1x1	1035.7	8
8	32	5.4	8	FeCo 2x1x1	192.4	8
16	32	6.8	16	FeCo 2x1x1	152.1	16
32	32	12.8	32	FeCo 2x1x1	80.9	32

Sign in to see the full dataset

Sep 29

ABACUS parallel layout: headline comparison

Data Views Docs Logs

Overview

22

Attribution

Connections

1

Referenced by

Ouro DFT was running at a fraction of its speed: one environment variable and a process layout make SCF 5–6× faster at a fifth of the cost
100 benchmark runs of ABACUS on Modal found OpenBLAS starting its own threads inside every process. One MPI process per core with k-point parallelism gives identical results 5–6× faster on the same 8-core workers. Shipped and validated on every CPU route.

Attribution

CC-BY-4.0 license

Best-of-two wall times for an Ouro DFT SCF (ABACUS v3.10.1, LCAO DZP, 100 Ry, kspacing 0.3, forces + stress) under the production layout and the fixed layouts, on Modal. relative_cost = (wall × cores) / production. All runs agree with production to ≤0.01 meV. Full per-run data: file 039a2753-a48b-4600-8a9d-02231f41afd1.

kpar	n_atoms	speedup	mpi_ranks	structure	best_wall_s	worker_cores
1	8	1	1	NiO	389.9	8
8	8	6.2	8	NiO	62.7	8
16	8	9.6	16	NiO	40.7	16
32	8	17.8	32	NiO	21.9	32
1	32	1	1	FeCo 2x1x1	1035.7	8
8	32	5.4	8	FeCo 2x1x1	192.4	8
16	32	6.8	16	FeCo 2x1x1	152.1	16
32	32	12.8	32	FeCo 2x1x1	80.9	32

Sign in to see the full dataset

2

Stats

Rows: 8
Columns: 11
Size: 32 KB
Downloads: 0

Join to comment

Connections

1

Referenced by

Ouro DFT was running at a fraction of its speed: one environment variable and a process layout make SCF 5–6× faster at a fifth of the cost
100 benchmark runs of ABACUS on Modal found OpenBLAS starting its own threads inside every process. One MPI process per core with k-point parallelism gives identical results 5–6× faster on the same 8-core workers. Shipped and validated on every CPU route.