Benchmarks

Measured evidence for the performance claims in this book. These are local ShortRun results, not a CI gate; hardware variance is real and the error columns are wide. Treat them as directional evidence, not as absolute numbers. For the same-model comparison against FluentValidation and DataAnnotations, see Peer benchmarks.

Machine profile

Item Value
OS Windows 11 (10.0.26300.9818)
CPU Intel Core i7-8850H 2.60 GHz (Coffee Lake), 1 CPU, 12 logical / 6 physical cores
Runtime .NET 10.0.14, X64 RyuJIT x86-64-v3
SDK .NET SDK 10.0.204
Tool BenchmarkDotNet 0.15.8 (historical)
Job ShortRun: 3 warmup, 3 measured iterations, 1 launch
Configuration Release, no debugger, pre-JIT

Historical measurement. These numbers were captured with the .NET 10 SDK. The repository now pins .NET 11 SDK 11.0.100 and C# 15 while the benchmark project targets the net10.0 baseline; re-run the suite before treating the numbers as a current performance claim.

Tooling. BenchmarkDotNet 0.15.8 does not recognize the net11.0 runtime moniker and throws GetRuntimeVersion not implemented for NotRecognized before running any benchmark. The repository now uses BenchmarkDotNet 0.16.0-preview.2, the first release with .NET 11 support. In 0.16 an Allocated value of - means zero managed allocations; older versions printed 0 B for the same result.

Reproduce with:

dotnet run -c Release --project src/Jeninnet.Validation.Benchmarks -- --filter '*ValidationBenchmarks*'
dotnet run -c Release --project src/Jeninnet.Validation.Benchmarks -- --filter '*LocalizationBenchmarks*'
dotnet run -c Release --project src/Jeninnet.Validation.Benchmarks -- --filter '*Base64ValidationBenchmarks*'

Validation engine

Models have 5, 10, and 30 validated properties. Generated uses real source-generated validators (the benchmark project references the analyzer); Fluent uses Validator<T>. Invalid models fail every rule; FailFast stops at the first error.

Method Mean Error StdDev Ratio Allocated Alloc Ratio
Fluent_Small_Valid 165.24 ns 152.51 ns 8.360 ns 1.00 0 B NA
Generated_Small_Valid 134.63 ns 186.36 ns 10.215 ns 0.82 0 B NA
Fluent_Small_Invalid 712.51 ns 716.75 ns 39.288 ns 4.32 968 B NA
Generated_Small_Invalid 676.21 ns 312.56 ns 17.132 ns 4.10 968 B NA
Fluent_Medium_Valid 147.78 ns 235.06 ns 12.884 ns 0.90 0 B NA
Generated_Medium_Valid 95.76 ns 32.42 ns 1.777 ns 0.58 0 B NA
Fluent_Medium_Invalid 1,358.55 ns 2,213.67 ns 121.339 ns 8.24 2,104 B NA
Generated_Medium_Invalid 1,397.88 ns 303.95 ns 16.660 ns 8.47 2,104 B NA
Fluent_Large_Valid 401.29 ns 230.56 ns 12.638 ns 2.43 0 B NA
Generated_Large_Valid 257.71 ns 227.69 ns 12.481 ns 1.56 0 B NA
Fluent_Large_Invalid 4,087.47 ns 4,375.20 ns 239.820 ns 24.78 6,024 B NA
Generated_Large_Invalid 3,871.51 ns 1,529.07 ns 83.813 ns 23.47 6,024 B NA
Fluent_Small_Invalid_FailFast 178.67 ns 88.25 ns 4.837 ns 1.08 208 B NA

Reading the numbers

  • Generated vs fluent: generated validators are faster on the valid path (135 vs 165 ns for 5 properties, 96 vs 148 ns for 10, 258 vs 401 ns for 30). Invalid paths are within run-to-run noise; error construction dominates.
  • Happy-path allocation: 0 B per operation. The ValidationErrorCollector and RootContextData are pooled and reused, and the rule-selection fast path returns the rule array without materialization. This confirms the zero-allocation happy-path contract.
  • Error path: allocations scale with the number of errors (968 B for 5 errors, 2,104 B for 10, 6,024 B for 30). Fail-fast cuts both time and allocation (179 ns / 208 B vs 713 ns / 968 B).
  • Benchmark-driven change: CompositeValidator now uses an indexed array fast path and a no-filter SelectRules return instead of foreach over IEnumerable plus materialization. Together these removed the residual 32 B/op measured during early benchmarking; the happy path is now zero. The allocation contract and stress evidence live in 07-allocation-audit.md.

Localization

Single built-in provider (DefaultValidationMessageProvider) with the built-in packs; the measurements below use English and Arabic, and the provider is the exactly-one-provider model used by the localization pipeline.

Method Mean Error StdDev Allocated
EnglishHit 26.24 ns 28.031 ns 1.536 ns 0 B
ArabicHit 28.13 ns 31.973 ns 1.753 ns 0 B
MissingKey 25.78 ns 4.394 ns 0.241 ns 0 B
MissingCulture 43.36 ns 16.249 ns 0.891 ns 0 B

Base64 rule

Method Mean Error StdDev Allocated
ValidBase64 80.70 ns 59.50 ns 3.262 ns 0 B

The Base64 rule and the engine's happy path are both allocation-free.

Peer comparison

A direct peer run is published in Peer benchmarks: on the repository machine the valid path measured 81.7 ns / 0 B for Jeninnet versus 226.7 ns / 632 B for FluentValidation. Treat those numbers as a machine-specific baseline, not a universal guarantee; run the benchmark harness on your own hardware before making capacity decisions. Generated and fluent validators share the same engine and allocation profile.

CI decision

Continuous performance-regression gating is not enabled. BenchmarkDotNet needs a quiet, pinned machine; hosted CI runners are too noisy for stable thresholds. The suite remains a manual evidence tool, run before releases and after engine changes that could shift numbers. A future self-hosted runner could add a budget gate; that is a separate decision.