Skip to main content

InfArray Benchmarks

Benchmarks comparing InfArray against a Luau table on the same workloads. Every workload stays under Luau's 2^26 per-table limit so the regular table is a fair baseline. These measure the overhead InfArray's chunking adds, not the cases where a table simply can't compete (i.e. once you cross 2^26 elements, the table stops being an option at all).

This page mirrors benchmark/README.md. The results table is regenerated by lute run benchmark/run.luau --md.

Running

From the repo root, with Lute on your PATH:

# Print results to the console
lute run benchmark/run.luau

# Re-run and rewrite the results table in this file
lute run benchmark/run.luau --md

# Only the benchmarks whose title matches a substring
lute run benchmark/run.luau --filter get

# Trade speed for accuracy (more timed samples)
lute run benchmark/run.luau --samples 10

# Quick pass at 10% of each workload size
lute run benchmark/run.luau --scale 0.1

Adding a benchmark

Copy an existing file in cases/ to cases/yourcase.bench.luau.
Fill in title, n, and the variants.
Add a require("./cases/yourcase.bench") line to the SPECS list in run.luau.

Interpreting the numbers

InfArray will always be slower per element than a native table. It does index math and an extra chunk lookup on every access. The point of the library is capacity, not speed. It holds more than 2^26 elements, which a single Luau table cannot. Use these benchmarks to keep that per-element overhead in check as the implementation changes, not to "beat" the table.

Results

CPU:      AMD Ryzen 9 9950X 16-Core Processor
OS:       Windows_NT (x86_64)
Date:     2026-06-23 11:49

PushBack - append N elements

Workload size: n = 4_194_304

Implementation	Time/op	Total	Throughput	vs baseline
InfArray	48.6 ns	203.96 ms	20.56 M/s	6.05× slower
Luau table	8.0 ns	33.71 ms	124.44 M/s	1.00× (baseline)

Set - overwrite N in-range elements

Workload size: n = 4_194_304

Implementation	Time/op	Total	Throughput	vs baseline
InfArray	31.4 ns	131.65 ms	31.86 M/s	9.72× slower
Luau table	3.2 ns	13.54 ms	309.67 M/s	1.00× (baseline)

Get - sequential read of N elements

Workload size: n = 4_194_304

InfArray uses the hoisted free function local Get = InfArray.Get.

Implementation	Time/op	Total	Throughput	vs baseline
InfArray	26.7 ns	111.92 ms	37.48 M/s	8.33× slower
Luau table	3.2 ns	13.44 ms	312.15 M/s	1.00× (baseline)

Get - random-access read of N elements

Workload size: n = 4_194_304

Implementation	Time/op	Total	Throughput	vs baseline
InfArray	104.4 ns	437.80 ms	9.58 M/s	2.85× slower
Luau table	36.7 ns	153.76 ms	27.28 M/s	1.00× (baseline)

Iterate - visit every present element

Workload size: n = 4_194_304

Iterate pays a callback per element

Implementation	Time/op	Total	Throughput	vs baseline
InfArray	13.7 ns	57.65 ms	72.75 M/s	5.23× slower
Luau table	2.6 ns	11.02 ms	380.59 M/s	1.00× (baseline)

`for index, value in arr` via __iter

Workload size: n = 4_194_304

__iter pays a closure call per element

Implementation	Time/op	Total	Throughput	vs baseline
InfArray	28.7 ns	120.31 ms	34.86 M/s	11.29× slower
Luau table	2.5 ns	10.66 ms	393.51 M/s	1.00× (baseline)

Find - linear search with needle at the end

Workload size: n = 4_194_304

Implementation	Time/op	Total	Throughput	vs baseline
InfArray	3.9 ns	16.17 ms	259.44 M/s	2.07× slower
Luau table	1.9 ns	7.83 ms	535.97 M/s	1.00× (baseline)

Results above are regenerated by lute run benchmark/run.luau --md. They are machine-specific. Treat them as relative ratios.

Running
Adding a benchmark
Interpreting the numbers
Results