Skip to content

Refresh benchmarks against current main #21

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
pscheid92 merged 1 commit into from
Jul 23, 2026
Merged

Refresh benchmarks against current main #21

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
100 changes: 53 additions & 47 deletions docs/BENCHMARKS.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,52 +1,56 @@
# Benchmarks

Measured on Apple M2 (8 cores, 16 GB), macOS 26, .NET 10, on 2026-07-23.
Numbers are refreshed against `main` at the current release (includes the
`IAeadCipher` backend seam and the pre-authentication parser bounds).

## CLI Comparison: age (Go) vs rage (Rust) vs age-sharp (C#/.NET)

Wall-clock time for encrypt and decrypt at various file sizes, averaged
over 5 iterations. Measured on Apple M2, .NET 10, AOT-compiled binary.
Includes process startup, key parsing, header processing, and I/O.
over 5 iterations. AOT-compiled binary. Includes process startup, key
parsing, header processing, and I/O.

All times in milliseconds (lower is better).

### Binary

| Size | Op | age (Go) | rage (Rust) | age-sharp (C#) |
|---|---|---:|---:|---:|
| 1 KB | enc | 21 ms | 20 ms | 22 ms |
| 1 KB | dec | 21 ms | 20 ms | 21 ms |
| 1 KB | enc | 21 ms | 21 ms | 22 ms |
| 1 KB | dec | 20 ms | 22 ms | 22 ms |
| 64 KB | enc | 21 ms | 21 ms | 22 ms |
| 64 KB | dec | 21 ms | 21 ms | 22 ms |
| 1 MB | enc | 22 ms | 23 ms | 24 ms |
| 1 MB | dec | 22 ms | 23 ms | 25 ms |
| 10 MB | enc | 34 ms | 47 ms | 41 ms |
| 10 MB | dec | 34 ms | 51 ms | 46 ms |
| 100 MB | enc | 155 ms | 285 ms | 211 ms |
| 100 MB | dec | 144 ms | 324 ms | 242 ms |
| 64 KB | dec | 20 ms | 21 ms | 22 ms |
| 1 MB | enc | 23 ms | 24 ms | 25 ms |
| 1 MB | dec | 22 ms | 24 ms | 25 ms |
| 10 MB | enc | 37 ms | 49 ms | 44 ms |
| 10 MB | dec | 35 ms | 52 ms | 47 ms |
| 100 MB | enc | 173 ms | 298 ms | 222 ms |
| 100 MB | dec | 147 ms | 323 ms | 244 ms |

### ASCII Armor (-a)

| Size | Op | age (Go) | rage (Rust) | age-sharp (C#) |
|---|---|---:|---:|---:|
| 1 KB | enc | 21 ms | 20 ms | 22 ms |
| 1 KB | dec | 21 ms | 20 ms | 21 ms |
| 1 KB | dec | 20 ms | 20 ms | 22 ms |
| 64 KB | enc | 21 ms | 21 ms | 22 ms |
| 64 KB | dec | 21 ms | 22 ms | 22 ms |
| 1 MB | enc | 25 ms | 24 ms | 24 ms |
| 64 KB | dec | 21 ms | 21 ms | 22 ms |
| 1 MB | enc | 27 ms | 24 ms | 25 ms |
| 1 MB | dec | 24 ms | 25 ms | 26 ms |
| 10 MB | enc | 67 ms | 55 ms | 45 ms |
| 10 MB | dec | 51 ms | 64 ms | 59 ms |
| 100 MB | enc | 436 ms | 359 ms | 251 ms |
| 100 MB | dec | 319 ms | 450 ms | 383 ms |
| 10 MB | enc | 72 ms | 56 ms | 46 ms |
| 10 MB | dec | 53 ms | 64 ms | 60 ms |
| 100 MB | enc | 456 ms | 366 ms | 275 ms |
| 100 MB | dec | 325 ms | 449 ms | 409 ms |

### Key Takeaways

- **Up to 1 MB**: All three implementations are within noise of each other
(~20-26 ms), dominated by process startup overhead.
- **Binary 100 MB**: Go still leads at 144-155 ms thanks to assembly-optimized
ChaCha20-Poly1305. AgeSharp (211-242 ms) beats rage (285-324 ms) after
switching to .NET's hardware-accelerated `ChaCha20Poly1305`.
- **Armored 100 MB encrypt**: AgeSharp (251 ms) now beats both rage (359 ms)
and Go (436 ms) after routing the push-encrypt path through `EncryptStream`
(~20-27 ms), dominated by process startup overhead.
- **Binary 100 MB**: Go still leads at 147-173 ms thanks to assembly-optimized
ChaCha20-Poly1305. AgeSharp (222-244 ms) beats rage (298-323 ms) using
.NET's hardware-accelerated `ChaCha20Poly1305`.
- **Armored 100 MB encrypt**: AgeSharp (275 ms) beats both rage (366 ms)
and Go (456 ms) — the push-encrypt path runs through `EncryptStream`
+ `ArmorStream` without any intermediate buffer.
- **Bounded memory**: all public push APIs (`Encrypt`, `Decrypt`,
`EncryptDetached`, `DecryptDetached`) stream chunk-by-chunk with pooled,
Expand Down Expand Up @@ -75,47 +79,49 @@ process startup overhead.

| Operation | 1 KB | 64 KB | 1 MB |
|---|---:|---:|---:|
| Encrypt | 98 us | 195 us | 2,038 us |
| Decrypt | 95 us | 206 us | 2,088 us |
| Encrypt (armored) | 99 us | 223 us | 2,349 us |
| Decrypt (armored) | 99 us | 281 us | 3,309 us |
| Encrypt | 102 us | 204 us | 2,092 us |
| Decrypt | 101 us | 215 us | 2,118 us |
| Encrypt (armored) | 102 us | 235 us | 2,485 us |
| Decrypt (armored) | 108 us | 294 us | 3,382 us |

Throughput at 1 MB: ~500 MB/s encrypt, ~490 MB/s decrypt.
Armored adds ~15-60% overhead due to Base64 encoding/decoding.
Throughput at 1 MB: ~500 MB/s encrypt, ~495 MB/s decrypt.
Armored adds ~15-60% overhead due to Base64 encoding/decoding. The
pre-auth parser bounds add one vectorized newline scan per buffered read
on the armored-decrypt path; the cost is within measurement noise.

Allocations at 1 KB are ~12 KB (Encrypt) / ~9 KB (Decrypt) — chunk
scratch buffers are rented from `ArrayPool<byte>.Shared` and reused
across operations. At 1 MB, most allocation is the output
`MemoryStream`'s growth, not the crypto path.
Allocations at 1 KB are ~20 KB (Encrypt) / ~17 KB (Decrypt) — a small,
constant per-operation cost (cipher setup plus chunk scratch buffers
rented from `ArrayPool<byte>.Shared` and reused across chunks). At 1 MB,
most allocation is the output `MemoryStream`'s growth, not the crypto path.

### Key Generation

| Operation | Time | Allocated |
|---|---:|---:|
| X25519 | 1,919 ns | 880 B |
| ML-KEM-768-X25519 | 246 ns | 88 B |
| X25519 | 1,982 ns | 880 B |
| ML-KEM-768-X25519 | 251 ns | 88 B |

### Recipient Wrap / Unwrap

| Operation | Time | Allocated |
|---|---:|---:|
| X25519 Wrap | 87 us | 3.6 KB |
| X25519 Unwrap | 84 us | 2.9 KB |
| ML-KEM-768-X25519 Wrap | 132 us | 46.7 KB |
| ML-KEM-768-X25519 Unwrap | 174 us | 69.9 KB |
| Scrypt Wrap | 1,785 us | 1,035 KB |
| Scrypt Unwrap | 1,778 us | 1,035 KB |

X25519 is the fastest at ~85 us. ML-KEM hybrid adds ~60-100% overhead
(still sub-200 us). Scrypt is intentionally slow (~1.8 ms) due to the
| X25519 Wrap | 89 us | 6.6 KB |
| X25519 Unwrap | 88 us | 5.8 KB |
| ML-KEM-768-X25519 Wrap | 137 us | 46.6 KB |
| ML-KEM-768-X25519 Unwrap | 182 us | 70.0 KB |
| Scrypt Wrap | 1,858 us | 1,035 KB |
| Scrypt Unwrap | 1,857 us | 1,035 KB |

X25519 is the fastest at ~88 us. ML-KEM hybrid adds ~55-105% overhead
(still sub-200 us). Scrypt is intentionally slow (~1.9 ms) due to the
password-hashing work factor.

### Random Access

| Operation | Time | Allocated |
|---|---:|---:|
| Sequential Read | 29.7 ms | 32.0 MB |
| Random Read | 31.5 ms | 34.2 MB |
| Sequential Read | 30.2 ms | 32.0 MB |
| Random Read | 32.1 ms | 34.2 MB |

Random reads are only ~6% slower than sequential thanks to the
chunk-based design.
Expand Down
Loading