lua-rs Performance History

Commit-keyed bench trajectory generated from harness/evidence/ledger.jsonl. Each point is one compare.sh run; the y-axis is the ratio of lua-rs to reference lua-c on the same workload. Parity at 1.00×.

Generated 2026-05-25T03:04:04+00:00
1046 measurements over 36 commits
Auto-refresh enabled

Wall-time geomean

1.27×

over 8 workloads at latest commit

RSS geomean

1.96×

resident-set vs reference lua-c

Best workload

0.38×

table_ops_long

Worst workload

2.07×

binarytrees

Wall-time ratio per workload

Ratio = lua-rs wall time ÷ reference lua-c wall time on the same workload. Lower is better; 1.00× is parity. Click a legend chip to mute that series. Off-screen points show as small triangles at the top edge.

y-max window

Tool: harness/bench/compare.sh · 3–5 runs per commit, min wall picked
Reference: PUC-Rio Lua 5.4.7 built from reference/lua-c/
Workloads: fibonacci · mandelbrot · binarytrees · closure_ops · table_ops · table_ops_long · string_ops · string_ops_long
Reading: Each point's y-value = lua-rs wall / lua-c wall on the matching workload, at the commit being benchmarked.

RSS ratio per workload

Ratio = lua-rs peak resident-set ÷ reference lua-c peak resident-set on the same workload. Lower is better; 1.00× is parity.

y-max window

Latest per-workload status

Workload	Wall ratio (latest)	RSS ratio (latest)	Parity gate	Last commit

External comparison — vs CppCXY/luars

One-off head-to-head (not part of the per-commit time series) against CppCXY/lua-rs — a separate Rust Lua runtime (Lua 5.5, with notable unsafe), raised in issue #12. Ratio = wall time ÷ reference C Lua 5.4.7; lower is better, <1.00× beats C. Output asserted byte-identical to reference C (drift flagged). Reproduce with harness/bench/compare_luars.sh.

Workload	C (s)	lua-rs (s)	luars (s)	lua-rs / C	luars / C	Note
fannkuch_redux	2.29	4.75	2.69	2.07×	1.17×
binary_trees	1.72	4.53	1.32	2.63×	0.77×	GC/alloc-heavy
nbody	1.77	3.72	1.99	2.10×	1.12×
spectral_norm	0.33	0.59	0.35	1.79×	1.06×
mandelbrot	0.52	0.93	0.56	1.79×	1.08×
partial_sums	0.18	0.31	0.20	1.72×	1.11×
fibonacci	2.39	4.45	2.31	1.86×	0.97×
binarytrees (own)	0.40	0.94	0.33	2.35×	0.82×	GC/alloc-heavy
closure_ops	0.17	0.34	0.17	2.00×	1.00×
mandelbrot_long	0.80	1.46	0.85	1.82×	1.06×
string_ops_long	0.78	0.99	0.86	1.27×	1.10×
table_ops_long	2.49	0.96	0.59	0.39×	0.24×	lua-rs & luars both beat C; luars output diverged*

Geomean vs C: lua-rs = 1.68× · luars = 0.90× (n=12)
Machine: Apple M3 Max · macOS arm64 · best-of-5 min wall · release builds
Reading: lua-rs leads only on table_ops (beats C); luars's edge concentrates on GC/alloc-heavy workloads (binary_trees), where lua-rs peak RSS is also markedly higher (402 MB vs 72 MB at trees depth 15).
Caveat: * luars output diverged on table_ops_long (likely a 5.5 vs 5.4 table semantic), so its number there is not fully trustworthy. macOS is the friendliest platform for both Rust impls; Linux/Windows typically shift ratios upward.

Recent runs

All bench measurements from harness/evidence/ledger.jsonl, newest first.

When	Commit	Subject	Workload	Wall ratio	RSS ratio	Runs