# Digest Benchmark - 2026-04-25

This benchmark compares Mercury `mhash` against common Windows and CLI hash tools on the same machine.

## Method

- Harness: `scripts/benchmark-mhash.ps1`
- Scenario: warm-cache CLI hashing; Windows does not provide a safe non-admin cache-drop primitive, so cold-cache disk results are intentionally excluded.
- Inputs: deterministic 1MiB, 64MiB, and 256MiB files under `target\mhash-benchmark\data`.
- Algorithms: `md5`, `sha1`, `sha256`, `sha512`.
- Repeats: 1 warmup plus 3 measured runs per tool/algorithm/size.
- Metrics: wall-clock time, throughput, process CPU time as percent of one core, and peak working set.
- Machine: `DESKTOP-EKBHO4B`, Windows `10.0.26200`, 32 logical processors.

## Tools

- Mercury `mhash 0.1.0`
- PowerShell `Get-FileHash` via PowerShell `7.6.1`
- Windows `certutil`
- uutils coreutils `0.8.0` commands: `md5sum`, `sha1sum`, `sha256sum`, `sha512sum`
- OpenSSL `3.6.1`

## Key Finding

The original Mercury `mhash` default `auto` mode memory-mapped large ordinary hash jobs. On this machine that increased peak working set to roughly the file size and was slower than streaming reads. The default was changed so streaming-friendly hash sets use one-pass buffered reads, while whole-file paths remain available for explicit `--io mmap` and for algorithms that benefit from whole-file/internal parallel processing.

## 256MiB Before/After

| Algorithm | Before auto ms | After auto ms | Speedup | Before peak MiB | After peak MiB | Peak drop MiB |
|---|---:|---:|---:|---:|---:|---:|
| md5 | 427.06 | 362.24 | 1.18x | 261.54 | 10.30 | 251.24 |
| sha1 | 201.81 | 163.24 | 1.24x | 255.72 | 10.55 | 245.17 |
| sha256 | 226.63 | 156.31 | 1.45x | 261.05 | 10.29 | 250.76 |
| sha512 | 341.78 | 273.64 | 1.25x | 264.60 | 10.57 | 254.03 |

## 256MiB Post-Optimization Comparison

| Algorithm | Tool | Mean ms | MiB/s | Peak WS MiB | CPU % of one core |
|---|---|---:|---:|---:|---:|
| md5 | openssl:dgst | 353.02 | 725.17 | 12.17 | 100.5 |
| md5 | mercury-mhash:auto | 362.24 | 706.71 | 10.30 | 94.9 |
| md5 | mercury-mhash:read | 364.38 | 702.57 | 10.30 | 95.8 |
| md5 | coreutils:md5sum | 377.49 | 678.17 | 9.44 | 96.6 |
| md5 | windows:certutil | 521.46 | 490.93 | 14.71 | 91.9 |
| md5 | powershell:Get-FileHash | 742.67 | 344.70 | 65.20 | 113.7 |
| sha1 | mercury-mhash:read | 155.19 | 1649.56 | 10.54 | 90.6 |
| sha1 | coreutils:sha1sum | 160.77 | 1592.36 | 9.45 | 90.7 |
| sha1 | mercury-mhash:auto | 163.24 | 1568.23 | 10.55 | 98.9 |
| sha1 | openssl:dgst | 196.64 | 1301.90 | 10.98 | 98.0 |
| sha1 | windows:certutil | 398.24 | 642.82 | 14.73 | 99.4 |
| sha1 | powershell:Get-FileHash | 641.07 | 399.33 | 66.26 | 114.5 |
| sha256 | mercury-mhash:read | 155.78 | 1643.35 | 10.29 | 93.6 |
| sha256 | mercury-mhash:auto | 156.31 | 1637.75 | 10.29 | 90.0 |
| sha256 | coreutils:sha256sum | 170.32 | 1503.08 | 10.36 | 97.9 |
| sha256 | openssl:dgst | 206.06 | 1242.36 | 10.99 | 91.0 |
| sha256 | windows:certutil | 270.40 | 946.74 | 14.75 | 92.5 |
| sha256 | powershell:Get-FileHash | 519.56 | 492.72 | 65.04 | 122.3 |
| sha512 | mercury-mhash:auto | 273.64 | 935.52 | 10.57 | 99.0 |
| sha512 | mercury-mhash:read | 273.88 | 934.70 | 10.31 | 93.2 |
| sha512 | openssl:dgst | 277.36 | 922.98 | 10.99 | 99.5 |
| sha512 | coreutils:sha512sum | 299.48 | 854.80 | 9.45 | 93.9 |
| sha512 | windows:certutil | 910.10 | 281.29 | 14.73 | 99.6 |
| sha512 | powershell:Get-FileHash | 1168.26 | 219.13 | 65.48 | 106.6 |

## Reproduce

```powershell
just mhash-benchmark
```

or:

```powershell
pwsh -NoProfile -ExecutionPolicy Bypass -File .\scripts\benchmark-mhash.ps1
```

The harness writes raw JSON, summary JSON, CSV, and Markdown reports under `target\mhash-benchmark\results`.