MD Trajectory Benchmarks

Trajectory benchmarks time frame-by-frame SASA over a whole trajectory, including reading the trajectory. All runs use zsasa 0.9.0, 128 sphere points, stride 1, the NACCESS classifier and explicit hydrogens.

Throughput summary#

Native zsasa occupies the top-left corner on every trajectory: the highest frame rate at the lowest peak memory. The Python integrations, zsasa + MDTraj and zsasa + MDAnalysis, reach a similar frame rate, and their memory is set by the Python trajectory loader.

Per-dataset comparison#

5wvo_C#

5wvo_C (1,001 frames, 3,858 atoms): 128 sphere points; zsasa at 10 threads.

6sup_A#

6sup_A (1,001 frames, 33,377 atoms): 128 sphere points; zsasa at 10 threads.

5vz0_A#

MDTraj was not measured on this trajectory.

5vz0_A (10,001 frames, 17,910 atoms): 128 sphere points; zsasa at 10 threads.

The "vs" columns are runtime ratios: the comparator's runtime divided by the row's runtime.

Beyond the core count#

The native trajectory path accepts more worker threads than the machine has cores. The benchmark machine has 10 logical CPUs.

Extra workers help a little here: 40 workers are 2% to 11% faster than 10, with the largest gain on the smallest system. Per-frame SASA values were identical at 10, 20 and 40 workers.

Workloads#

Memory interpretation#

zsasa streams trajectory frames and keeps memory close to the current-frame working set. In contrast, the mdsasa-bolt path uses an MDAnalysis front-end that materializes atom data for every frame before the Rust SASA core runs, so its peak memory grows with trajectory length.

Validation pointer#

Agreement with MDTraj on 5wvo_C is covered in SASA Validation.

Evidence source#

The charts and tables on this page are generated from results/tables/md_summary.csv and md_thread_scaling.csv in N283T/zsasa-benchmarks.