Single-File Stress Benchmarks

Single-file benchmarks focus on large structures and parser-heavy cases that are hidden by directory-level averages. The pinned current benchmark suite uses eight protein-only PDB inputs, 100 sphere points, and 10 threads for the headline results.

Runtime and memory

Figure 1. Single-file runtime bars. Absolute wall-clock time across the eight curated structures at 10 threads.

Figure 2. Single-file peak RSS bars. Peak memory remains lower for zsasa across the large-structure and parser-stress cases.

Figure 3. Single-file runtime speedup ratios. The n× view highlights parser and very-large-structure stress cases.

Figure 4. Single-file RSS reduction ratios. Memory ratios are shown alongside absolute RSS bars.

Selected 10-thread results:

Structure	Mode	Runtime	RSS	Speedup vs FreeSASA	Speedup vs RustSASA
AF-Q6ZS30-F1, 21,611 atoms	zsasa f64	0.026 s	9.1 MiB	1.99×	1.69×
AF-Q6ZS30-F1, 21,611 atoms	zsasa bitmask f64	0.034 s	11.4 MiB	1.54×	1.31×
5vyc, 249,168 atoms	zsasa f64	0.238 s	96.8 MiB	2.26×	52.3×
5vyc, 249,168 atoms	zsasa bitmask f64	0.197 s	96.8 MiB	2.73×	63.0×
8rbs, 164,605 atoms	zsasa f64	0.118 s	48.1 MiB	119.7×	1.97×
8rbs, 164,605 atoms	zsasa bitmask f64	0.104 s	51.9 MiB	136.5×	2.25×
9fqr, 4,506,416 atoms	zsasa f64	4.696 s	1,615 MiB	40.9×	1.86×
9fqr, 4,506,416 atoms	zsasa bitmask f64	3.788 s	1,616 MiB	50.6×	2.30×

Dataset

Inputs were normalized to protein-only PDB files for fair comparator runs: hydrogens, alternate conformations, ligands, waters, and non-L-peptide chains were removed; atom and residue identifiers were wrapped into PDB field limits when needed.

Structure	Role	Atoms	Chains	Note
AF-P49792-F10	single-medium	10,919	1	AFDB single-chain case
AF-Q6ZS30-F1	single-large	21,611	1	AFDB large single-chain case
AF-0000000066638622	multi-medium	14,618	2	AFDB-derived two-chain case
AF-0000000065781219	multi-large	24,140	2	AFDB-derived two-chain case
3jc8	100k-atom	107,500	3	PDB assembly
5vyc	RustSASA stress	249,168	4	Parser-heavy RustSASA case
8rbs	FreeSASA stress	164,605	5	PDB coordinate-overflow parser case
9fqr	maximum-size	4,506,416	57	Largest assembly in this subset

Lahuta is excluded from this suite because the benchmarked SASA command targets AlphaFold-style chain-A inputs and cannot process this mixed multi-chain subset.

Parse versus SASA timing

zsasa records component timings with its native --timing path. The benchmark harness also used timing-enabled comparator builds so parse and kernel effects could be separated.

Figure 5. Parse time versus SASA-kernel time. On the largest assembly, zsasa f64 spent about 708 ms parsing and 3,754 ms in SASA calculation; bitmask reduced the SASA component to about 2,861 ms.

Selected component timings:

Structure	Tool/mode	Parse	SASA kernel	Wall runtime
5vyc	zsasa f64	40.4 ms	184.1 ms	0.238 s
5vyc	RustSASA	12,204.5 ms	349.9 ms	12.430 s
8rbs	zsasa f64	23.5 ms	84.3 ms	0.118 s
8rbs	FreeSASA	119.4 ms	13,467.4 ms	14.150 s
9fqr	zsasa f64	707.8 ms	3,753.7 ms	4.696 s
9fqr	FreeSASA	3,486.1 ms	167,332.8 ms	191.876 s
9fqr	RustSASA	1,526.6 ms	7,100.0 ms	8.731 s

Thread scaling

Figure 6. Runtime versus thread count across the curated single-file subset. Thread scaling depends on structure size and parser behavior; the largest structures benefit most from parallel SASA calculation.

Caveats

• The 8rbs FreeSASA result primarily exposes PDB fixed-width coordinate overflow and coordinate misreading rather than ordinary FreeSASA kernel scaling.
• The 5vyc RustSASA result is dominated by parser time.
• zsasa accepted the raw large-structure inputs directly, but all tools were timed on the same normalized PDB files for fairness.

Evidence source

The values above are exported from zsasa-benchmarks/results/tables/single_file_t10_summary.csv and single_file_thread_scaling.csv.