MD Trajectory SASA Benchmarks

Performance comparison of SASA calculation on molecular dynamics trajectories.

Note: Times include trajectory loading and SASA calculation. All tools use 100 test points per atom (Shrake-Rupley algorithm).

TL;DR

Dataset	Atoms	Frames	zsasa CLI bitmask (f32, 10t)	vs MDTraj	vs mdsasa-bolt	RSS
5wvo_C	3,858	1,001	0.78s	30x	5.8x	17 MB
6sup_A	33,377	1,001	6.92s	132x	8.5x	113 MB
5vz0_A	17,910	10,001	38.04s	>2h (timeout)	>2h (timeout)	63 MB

• zsasa CLI bitmask is the fastest variant: ~2x faster than standard zsasa on large systems
• Memory is constant with frame count (streaming I/O) — 63 MB for 10K frames vs Python tools using 3–6 GB
• MDTraj and mdsasa-bolt timed out (>2h) on the 10K-frame dataset that zsasa completes in 38 seconds

Test Environment

Item	Value
Machine	MacBook Pro
Chip	Apple M4 (10 cores: 4P + 6E)
Memory	32 GB
OS	macOS 15.3.2 (Darwin 24.6.0)

Implementations

Tool	Language	Threading	I/O	Notes
zsasa CLI	Zig	Configurable	Native XTC reader	f32/f64 precision, streams frames
zsasa CLI bitmask	Zig	Configurable	Native XTC reader	LUT bitmask, streams frames
zsasa.mdtraj	Python/Zig	Configurable	MDTraj (C)	MDTraj loads trajectory, zsasa computes SASA
zsasa.mdtraj bitmask	Python/Zig	Configurable	MDTraj (C)	LUT bitmask variant
zsasa.mdanalysis	Python/Zig	Configurable	MDAnalysis	MDAnalysis loads trajectory, zsasa computes SASA
zsasa.mdanalysis bitmask	Python/Zig	Configurable	MDAnalysis	LUT bitmask variant
MDTraj	Python/C	Single-threaded	MDTraj (C)	shrake_rupley, no parallel option
mdsasa-bolt	Rust	All cores (rayon)	MDAnalysis/Python	RustSASA via Python bridge

Results

6sup_A_analysis (Large: 33,377 atoms, 1,001 frames)

Benchmark: warmup=1, runs=3, threads=1,8,10.

10-Thread Comparison

Tool	Time (s)	FPS	vs MDTraj	vs mdsasa-bolt	RSS
zsasa CLI (f32, bitmask)	6.92	145	132.4x	8.5x	113 MB
zsasa CLI (f64, bitmask)	7.16	140	127.9x	8.2x	120 MB
zsasa.mdtraj bitmask	7.31	137	125.3x	8.0x	1.5 GB
zsasa.mdanalysis bitmask	7.71	130	118.8x	7.6x	988 MB
zsasa CLI (f32)	14.79	68	61.9x	4.0x	111 MB
zsasa.mdtraj	15.59	64	58.7x	3.8x	1.6 GB
zsasa CLI (f64)	15.79	63	58.0x	3.7x	116 MB
zsasa.mdanalysis	17.11	59	53.5x	3.4x	1.0 GB
mdsasa-bolt	58.53	17	15.6x	baseline	10.9 GB
MDTraj	915.57	1.1	baseline	—	874 MB

Key findings:

• zsasa CLI bitmask (f32) is 132x faster than MDTraj, 8.5x faster than mdsasa-bolt
• Bitmask variants are ~2.1x faster than standard zsasa (6.92s vs 14.79s)
• mdsasa-bolt uses 10.9 GB RAM vs zsasa CLI's 113 MB (96x difference)

Tool Comparison (10t)	Memory

5vz0_A_protein (Medium: 17,910 atoms, 10,001 frames)

Benchmark: warmup=1, runs=3, threads=1,8,10.

MDTraj and mdsasa-bolt timed out at the 2-hour limit (--timeout 7200). MDTraj is too slow for 10K frames of this size; mdsasa-bolt likely exhausted memory (see mdsasa-bolt Performance).

10-Thread Comparison

Tool	Time (s)	FPS	RSS
zsasa.mdtraj bitmask	37.41	267	5.8 GB
zsasa.mdanalysis bitmask	37.77	265	3.2 GB
zsasa CLI (f32, bitmask)	38.04	263	63 MB
zsasa CLI (f64, bitmask)	38.99	256	67 MB
zsasa.mdtraj	81.55	123	5.8 GB
zsasa CLI (f32)	82.58	121	61 MB
zsasa.mdanalysis	82.69	121	3.1 GB
zsasa CLI (f64)	86.08	116	64 MB

Key findings:

• Bitmask variants ~2.1x faster than standard (38s vs 83s)
• zsasa CLI processes 10,001 frames in 38 seconds (263 FPS) with only 63 MB RAM
• Python bindings slightly faster than CLI on bitmask (trajectory pre-loaded vs streaming), but use 50–90x more memory

Tool Comparison (10t)	Memory

5wvo_C_analysis (Small: 3,858 atoms, 1,001 frames)

Benchmark: warmup=1, runs=3, threads=1,8,10.

10-Thread Comparison

Tool	Time (s)	FPS	vs MDTraj	vs mdsasa-bolt	RSS
zsasa CLI (f32, bitmask)	0.78	1,291	30.0x	5.8x	17 MB
zsasa CLI (f64, bitmask)	0.79	1,263	29.4x	5.6x	18 MB
zsasa.mdtraj bitmask	0.88	1,144	26.6x	5.1x	210 MB
zsasa.mdanalysis bitmask	1.20	838	19.5x	3.7x	172 MB
zsasa.mdtraj	1.64	612	14.2x	2.7x	195 MB
zsasa CLI (f32)	1.69	593	13.8x	2.6x	14 MB
zsasa CLI (f64)	1.77	568	13.2x	2.5x	15 MB
zsasa.mdanalysis	1.99	505	11.7x	2.2x	166 MB
mdsasa-bolt	4.47	224	5.2x	baseline	1.4 GB
MDTraj	23.29	43	baseline	—	156 MB

Key findings:

• On small systems, bitmask advantage is smaller (~2.2x vs ~2.1x on large)
• zsasa CLI bitmask processes 1,001 frames in under 1 second at 1,291 FPS
• Even on a small system, mdsasa-bolt uses 1.4 GB vs zsasa CLI's 17 MB

Tool Comparison (10t)	Memory

Memory Usage

Tool	5wvo_C (4k atoms, 1k fr)	5vz0_A (18k atoms, 10k fr)	6sup_A (33k atoms, 1k fr)
zsasa CLI bitmask	17 MB	63 MB	113 MB
zsasa CLI	14 MB	61 MB	111 MB
zsasa.mdanalysis bitmask	172 MB	3.2 GB	988 MB
zsasa.mdtraj bitmask	210 MB	5.8 GB	1.5 GB
MDTraj	156 MB	—	874 MB
mdsasa-bolt	1.4 GB	—	10.9 GB

• zsasa CLI memory scales with atom count, not frame count (streaming I/O)
• Python tools load the entire trajectory into memory — scales linearly with frames
• For 5vz0 (10k frames, 18k atoms): zsasa CLI uses 63 MB while Python bindings use 3–6 GB

Bitmask Variants

The _bitmask variants use LUT bitmask neighbor lists:

• ~2.1x faster than standard zsasa across all system sizes
• Minimal memory overhead for CLI variants (~2 MB extra)
• Python bitmask bindings are faster but still load full trajectories into memory
• On small systems (5wvo, 4k atoms), bitmask advantage is ~2.2x; on large systems (6sup, 33k atoms), ~2.1x

mdsasa-bolt Performance

mdsasa-bolt (RustSASA via Python) shows high memory usage across all datasets:

Dataset	mdsasa-bolt RSS	zsasa CLI RSS	Ratio
5wvo_C (4k atoms)	1.4 GB	17 MB	82x
6sup_A (33k atoms)	10.9 GB	113 MB	96x

On the previous 10k-frame dataset (6sup_R1), mdsasa-bolt showed severe performance degradation (147x slower than zsasa), becoming slower than single-threaded MDTraj. This pattern is consistent with memory-bound scaling issues.

SASA Validation

Per-frame total SASA compared against MDTraj shrake_rupley as baseline, using 5wvo_C (3,858 atoms, 1,001 frames).

Note on baseline: MDTraj and zsasa use different sphere point placement (MDTraj uses reversed winding order). This causes small systematic differences that diminish with increasing test points. Bitmask error is independent of point count due to the LUT approximation. Plots use zoomed axes — the spread appears larger than it is in practice.

R² vs MDTraj by Test Points

Tool	100 pts	200 pts	500 pts	1000 pts
zsasa CLI (f64)	0.8716	0.9664	0.9938	0.9983
zsasa CLI (f32)	0.8716	0.9664	0.9938	0.9983
zsasa.mdtraj	0.8716	0.9664	0.9938	0.9983
zsasa.mdanalysis	0.9724	0.9885	0.9718	0.9602
zsasa CLI bitmask (f64)	0.3980	0.6437	0.7537	0.7916
zsasa CLI bitmask (f32)	0.3981	0.6438	0.7537	0.7917

• Standard variants converge rapidly: R² > 0.99 at 500 points, > 0.998 at 1000 points
• Bitmask variants plateau around R² ~0.79 regardless of point count (LUT approximation error)
• f32 and f64 produce virtually identical results
• zsasa CLI f64 vs zsasa.mdtraj: R² = 1.000 (identical SASA engine, different I/O path — max error 0.008%)

ΔR² (Frame-to-Frame SASA Changes)

ΔR² measures how well frame-to-frame SASA changes (ΔSASA) correlate — which is what matters most in MD analysis.

Tool	100 pts	200 pts	500 pts	1000 pts
zsasa CLI (f64)	0.8918	0.9527	0.9895	0.9962
zsasa CLI (f32)	0.8918	0.9527	0.9895	0.9962
zsasa CLI bitmask (f64)	0.8902	0.9501	0.9872	0.9941
zsasa CLI bitmask (f32)	0.8902	0.9502	0.9874	0.9941

• Bitmask ΔR² is nearly identical to standard — the systematic offset from LUT cancels out in frame differences
• At 100 points: bitmask ΔR² = 0.890 vs standard ΔR² = 0.892 (negligible gap)
• At 1000 points: both converge to ΔR² > 0.994

Practical Significance

The absolute R² values for bitmask may look low (~0.79), but the scatter plots use zoomed axes that exaggerate the spread. The actual max per-frame error is 1.5–3%, and more importantly, the ΔR² shows that frame-to-frame trends are tracked with >0.99 correlation at 500+ points — even for bitmask.

For MD trajectory analysis where SASA trends and changes matter more than absolute values, bitmask accuracy is sufficient. If precise absolute values are needed, use the standard f64 variant.

XTC reader consistency	Bitmask comparison

Running Benchmarks

# Run benchmark
./benchmarks/scripts/bench_md.py \
    --xtc trajectory.xtc \
    --pdb topology.pdb \
    --name my_bench \
    --threads "1,8,10"

# Specific tools only
./benchmarks/scripts/bench_md.py \
    --xtc trajectory.xtc \
    --pdb topology.pdb \
    --name my_bench \
    --tool zig --tool zig_bitmask --tool zsasa_mdtraj

# Analysis
./benchmarks/scripts/analyze_md.py all --name my_bench
./benchmarks/scripts/analyze_md.py summary --name my_bench
./benchmarks/scripts/analyze_md.py bar --name my_bench
./benchmarks/scripts/analyze_md.py memory --name my_bench

Options

Option	Description	Default
--xtc	XTC trajectory file	(required)
--pdb	Topology PDB file	(required)
--name, -n	Benchmark name	(required)
--threads, -T	Thread counts: 1,4,8 or 1-10	1,8
--runs, -r	Number of benchmark runs	3
--warmup, -w	Number of warmup runs	1
--stride, -s	Frame stride	1
--n-points	Test points per atom	100
--tool, -t	Tool: zig, zig_bitmask, zsasa_mdtraj, zsasa_mdtraj_bitmask, zsasa_mdanalysis, zsasa_mdanalysis_bitmask, mdtraj, mdsasa_bolt (repeatable)	all
--output, -o	Output directory	results/md/<n_points>/<name>
--dry-run	Show commands without running	false

Notes

• f32 and f64 precision produce nearly identical timing (~3–5% difference)
• SASA accuracy and XTC reader consistency validated in validation.md
• mdsasa-bolt "all" = 10 threads on this system

Related Documents

• Single-File Benchmarks — Per-structure performance across 2,013 structures
• Batch Processing Benchmarks — Proteome-scale throughput
• SASA Validation — Accuracy validation against FreeSASA