Sep 29

ABACUS MPI/OpenMP layout benchmark: all runs

Data Docs Logs

Overview

00

Attribution

Connections

1

Linked from

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

Every run from the Ouro DFT (ABACUS v3.10.1, LCAO DZP, 100 Ry, kspacing 0.3, SCF + forces/stress) parallel-layout benchmark on Modal: layout (worker:ranks x threads[:kpar[:binding]]), OpenBLAS thread setting, wall time, energy, moment, ABACUS per-section timings, and host CPU family. Four series: first layout matrix, timing + pinning, OpenBLAS pinned, production baseline.

0

Stats

Size: 10.84 KB
Type: .csv file
Downloads: 0

Join to comment

Connections

1

Linked from

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.