Rings on one thread

primitive::ringNetwork against OVITO 3.17.2 FindRingsModifier. The cutoff graph is built before the timer on both sides. The cutoff is 3.5 Å and the largest ring has 6 members. Eight timed calls follow two warmups. The median of those eight is the mean of the two central samples.

Frames

These are the three sizes in the earlier one-thread campaign.

name

molecules

frame

tri-136

136

genice_sH.lammpstrj, guest-free structure H

orth-4k

4096

mW_cubic.lammpstrj, cubic ice

orth-1e5

110592

that cubic cell, replicated 3 by 3 by 3

The per-size census is the same in both codes.

molecules

4

5

6

136

12

120

68

4096

0

0

8192

110592

0

0

221184

This session

Eight cores, one thread per core, Intel Xeon, family 6, model 207, stepping 2. GCC 15.2.0, release build, OpenMP. OVITO 3.17.2. The core count is 8. The record does not name the host.

molecules

threads

d-SEAMS (ms)

OVITO (ms)

136

1

0.212

0.550

4096

1

2.715

13.548

110592

1

82.343

474.104

136

8

0.090

0.830

4096

8

1.521

13.946

110592

8

27.246

298.562

On one thread d-SEAMS is faster at each of the three sizes. Samples are in markdown/rings-one-thread.json.

Earlier campaign

That record used another 8-core Xeon and an older build of the search. OVITO was faster there. The frames, the cutoff, the maximum ring size, and the timing boundary (graph outside the timer) are the ones above.

molecules

d-SEAMS (ms)

OVITO (ms)

136

1.078

0.697

4096

31.617

15.831

110592

864.701

511.024

Rerun

meson setup build --buildtype=release -Dwith_tests=true \
  -Dwith_mpi=disabled -Dwith_openmp_offload=disabled
ninja -C build -j2 tests/bench_rings_h2h
python3 scripts/rings_one_thread.py \
  --bench build/tests/bench_rings_h2h \
  --json markdown/rings-one-thread.json

The interpreter must provide OVITO 3.17.2. Set RINGS_PYTHON when that is not python3. Snakemake rule rings_one_thread runs the same driver. Airspeed Velocity class RingsOneThread times one search on each side after the graph has been built.