Linkcell strong scaling

An orthorhombic frame is cut by linkcell’s weighted recursive coordinate bisection of the periodic cell grid. The balance bound is the tightest a cell-aligned cut can promise: one plus the part count times the heaviest cell, over the total weight. The first call of an analysis weights cells by how many points they hold. A later call of that analysis weights them by the work measured on the first call.

The halo is painted in the same walk that visits the cells a rank owns. A point in an owned cell stays owned. A point in a painted cell that another rank owns is a ghost. The same cut is used inside one process. Each thread owns a block of cells and writes the neighbour rows of the atoms in those cells. The halo is painted in that owned walk, so the threads do not share one cell mesh. The cut is reused on the next call with the same grid. Neighbour rows of an unchanged frame are refilled in place.

A skewed cell does not use that mesh. It keeps the slab, pencil, or block cut.

Session

One sitting, 12 November 2026. host_cores was read with OMP_NUM_THREADS unset and reported 8. The compiler is g++ 15.2.0. The library is a Meson release build. Eight samples follow one warm-up. The median of an even count is the mean of the two central samples. Speedup is against one thread. The four operations are the neighbour rows, CHILL+, the Steinhardt q6, and primitive rings of depth 6. The ice is eight molecules in a cubic cell. No host name is recorded.

On 884736 molecules, eight threads build the neighbour rows in 25.7 ms, 5.72 times the one-thread time. CHILL+, q6 and the ring census use those rows. OVITO 3.17.2 CreateBonds and freud 3.6.1 AABBQuery were timed on the same ice in this sitting. OVITO stores one record per undirected bond. freud stores both directions, which is the d-SEAMS pair count.

OpenMP, 110592 molecules

threads

neigh ms

speedup

chill ms

speedup

q6 ms

speedup

rings ms

speedup

1

15.2

1.00

47.4

1.00

43.5

1.00

83.4

1.00

2

11.6

1.32

34.8

1.36

26.4

1.65

64.0

1.30

4

7.4

2.05

28.3

1.67

16.7

2.60

34.0

2.46

8

5.6

2.71

23.8

1.99

10.7

4.07

20.2

4.13

OpenMP, 884736 molecules

threads

neigh ms

speedup

chill ms

speedup

q6 ms

speedup

rings ms

speedup

1

146.9

1.00

442.6

1.00

497.9

1.00

666.9

1.00

2

74.2

1.98

285.8

1.55

318.5

1.56

406.4

1.64

4

41.4

3.55

224.2

1.97

225.7

2.21

236.4

2.82

8

25.7

5.72

157.8

2.81

181.0

2.75

158.9

4.20

Neighbours against OVITO and freud

Same frames and the same sitting. Times are milliseconds.

molecules

threads

d-SEAMS

OVITO 3.17.2

freud 3.6.1

110592

1

15.2

43.9

101.7

110592

8

5.6

11.2

103.2

884736

1

146.9

302.9

1020.0

884736

8

25.7

104.1

1023.9

MPI, one thread per rank

These ranks were timed in the earlier sitting the same day, before each thread owned a block. Speedup is against that sitting’s one rank.

molecules

ranks

neigh ms

speedup

chill ms

speedup

q6 ms

speedup

rings ms

speedup

110592

1

15.4

1.00

33.2

1.00

39.8

1.00

75.1

1.00

110592

2

18.7

0.83

29.6

1.12

32.9

1.21

39.9

1.88

110592

4

10.3

1.49

17.3

1.92

24.0

1.66

21.4

3.51

110592

8

4.1

3.76

7.6

4.38

8.1

4.91

10.7

7.02

884736

1

166.4

1.00

340.2

1.00

556.7

1.00

618.4

1.00

884736

2

127.8

1.30

238.0

1.43

328.2

1.70

336.3

1.84

884736

4

62.2

2.68

115.9

2.94

156.7

3.55

168.0

3.68

884736

8

26.4

6.30

59.3

5.74

67.8

8.22

84.6

7.31