Benchmarks

Time a 2.14.0 build and write down the run. The guest-code medians are on Literature head-to-head. The thread width each guest receives is on Reproduce the literature head-to-head. The pinned neighbour session is Cutoff neighbour lists.

Build

pixi run build configures bbdir with -Dwith_tests=true and compiles it. The same two steps by hand:

meson setup bbdir --prefix=$CONDA_PREFIX -Dwith_tests=true
ninja -C bbdir

The timers are bbdir/tests/bench_*. pixi run test runs the test suite and does not run them.

A host-only build takes the host branch of the device code. SEAMS_OFFLOAD=0 forces that path. SEAMS_OFFLOAD=1 selects the offload path when the build has one. Record the host, the compiler, the commit, the thread count, the cpu affinity, and SEAMS_OFFLOAD.

The comparisons below use the Python scripts under scripts/.

Neighbour list

bench_cutoff_list times nneigh::orthoFullCsr and neighListO on ideal ice Ic. The lattice constant is 6.35 Å and the cutoff is 3.5 Å. Molecules must be 8 n^3. The arguments are molecules, threads, and repetitions. The defaults are 110592, 1, and 8. Every repetition is a sample. The printed median is the sample at index n/2 after sorting.

bbdir/tests/bench_cutoff_list 110592 1 8

scripts/bench_neighbours_threads.py runs that binary and the same list in vesin, freud, and matscipy, at 10648, 110592, and 884736 molecules and at 1, 2, 4, and 8 threads. Set NEIGH_BENCH to the binary. The script’s own default is build-release/tests/bench_cutoff_list. --reps defaults to 8. --out writes the lines.

Triclinic list, CHILL+, rings, and diamond

bench_v2 takes a mode, a molecule count, a thread count, and a repetition count. Molecules must be 8 n^3. tilt builds the triclinic cell, with xy = 0.25 L, xz = 0.05 L, and yz = 0.12 L. Any other mode uses the orthorhombic cell. The cutoff is 3.5 Å and the neighbour build is inside the timer. The first call is the warm-up and is dropped. The printed median is again the sample at index n/2 after sorting.

mode

what the timer includes

tilt

the triclinic cutoff list

chill

that list, then CHILL+

rings

that list, then primitive rings to depth 6

symmetry

that list, then the diamond common-neighbour label

bbdir/tests/bench_v2 tilt 110592 1 8

rings is a full search of one frame. The carried update is bench_hop_stride, below. scripts/bench_neighbours_v2.py drives bench_v2 and the matching vesin, freud, mdapy, and OVITO calls. Set NEIGH_BENCH. The script’s default is bbdir-release/tests/bench_v2. An output path, when given, is the first argument. NEIGH_ONLY limits the matrix to mode:molecules:threads.

Local q4 and q6

bench_ql reads a frame: a little-endian uint32 count, a float64 box length, then n times three float64 coordinates. The arguments are the frame, the degree, and the number of timed calls. One warm-up runs first. The neighbour build, at cutoff 3.5 Å, is inside every sample. A last path writes the warm-up \(q_l\) as raw float64.

bbdir/tests/bench_ql frame.bin 6 8

scripts/bench_ql_threads.py builds ideal ice Ic at bond 2.75 Å and one shifted liquid, and times degrees 4 and 6 against OVITO and freud. --out is required. The binary is SEAMS_QL_BENCH, or bbdir-release/tests/bench_ql, or bbdir/tests/bench_ql. One warm-up, then eight samples. The median of an even count is the mean of the two central values, and the interquartile range is Hyndman-Fan type 7.

Primitive rings against OVITO

bench_rings_h2h times the ring search on a cutoff graph built before the timer. --repl reads commands: go repeats that search, in times the neighbour list, the index, and the search together, and quit stops.

bbdir/tests/bench_rings_h2h input/traj/exampleTraj.lammpstrj 1 2 3.5 6 8 2

The arguments after the dump are frame, atom type, cutoff, maximum ring, repetitions, and warm-up. The defaults for the last two are 8 and 2. Type 2 is the oxygen in exampleTraj.lammpstrj.

scripts/rings_one_thread.py runs the binary and OVITO CreateBondsModifier plus FindRingsModifier at one thread and at eight. --bench is the binary and --json is the output path.

Cage stage

Run bench_cages from input/. The default dump is traj/mW_cubic.lammpstrj, frame 1, three repetitions, atom type 1. The later arguments are the HC template, the DDC template, and the atom type. The templates in the tree are templates/hc.xyz and templates/ddc.xyz.

cd input && ../bbdir/tests/bench_cages

Each printed time is the fastest repetition. The lines are the ring network to depth 7, findHC, findDDC, both of those, local \(q_6\), and, when the templates load, the shape match. A nauty certificate of the six-rings is included when nauty is linked.

bench_ddc_ladder times findDDC alone. Neighbours, rings, and findHC run before the timer. The arguments are the dump, the warm-up count, and the repetition count.

bbdir/tests/bench_ddc_ladder input/traj/mW_cubic.lammpstrj 1 8

Incremental updates

bench_churn_sweep moves a ball of molecules from frame A to frame B and times RingUpdater and AffiliationUpdater against a full pass. Each printed time is the median of the repetitions. Every incremental result is checked against the batch result.

bbdir/tests/bench_churn_sweep input/traj/mW_cubic.lammpstrj

The optional arguments are frame A, frame B, cutoff, atom type, repetitions, a ring full-pass fraction, and an affiliation batch fraction.

bench_trajectory_incremental walks a dump. RingUpdater is checked against ringNetwork, and AffiliationUpdater against the batch affiliation. The graph is cutoff, knn, or knn-union. The default graph is knn.

bbdir/tests/bench_trajectory_incremental input/traj/mW_cubic.lammpstrj

The optional arguments are the frame count, the atom type, the maximum ring depth, the skin, and the graph.

bench_hop_stride times the carried primitive-ring update against a full search of the same frame. The graphs are the 3.5 Å cutoff list and the seeded pair used by seams cages: mutual and union 4-nearest, candidate cutoff 5 Å, rings through depth 6. The names in the log are cutoff, seeded-strict, seeded-union, and seeded. One warm-up, then eight repetitions, unless --reps and --warmup say otherwise. The printed carry and rebuild times are type-7 medians. The printed speedup is the sum of the per-frame median rebuild times divided by the sum of the per-frame median carry times. A transition is inside the ball when the update enumerates fewer sources than the molecule count.

bbdir/tests/bench_hop_stride time input/traj/mW_cubic.lammpstrj --graph both

crossover uses --frame-a and --frame-b. --type selects the atom type. The carried-ring definition is on Linkcell neighbours and carried updates.

Device batch

bench_gpu_batch reads a dump and prints resident yes or resident no. A host build that cannot keep the frames on the device stops there. When they stay resident, the log has the cold upload, compute, and download, then the warm repeat with the smallest compute time.

bbdir/tests/bench_gpu_batch input/traj/mW_cubic.lammpstrj

The optional arguments are the frame count, the atom type, and the number of warm repeats. The defaults are 11, 1, and 5.