benchmarkdotnet-setup

BenchmarkDotNet Setup

Write correct BenchmarkDotNet benchmarks (no common pitfalls) and read the results properly — so "which is faster" is answered with data, not vibes.

Input — point it at your code

Works on a target: a file, folder, project, or GitHub URL. Identify the methods/implementations to compare; if the user pasted results, go straight to interpretation.

Project + benchmark class

dotnet add package BenchmarkDotNet

[MemoryDiagnoser]                     // allocations matter as much as time
[Orderer(SummaryOrderPolicy.FastestToSlowest)]
public class StringJoinBenchmarks
{
    private readonly string[] _items = Enumerable.Range(0, 1000).Select(i => i.ToString()).ToArray();

    [Benchmark(Baseline = true)]
    public string StringConcat() { var s = ""; foreach (var i in _items) s += i; return s; }

    [Benchmark]
    public string StringBuilder_() { var sb = new StringBuilder(); foreach (var i in _items) sb.Append(i); return sb.ToString(); }

    [Benchmark]
    public string StringJoin() => string.Join("", _items);
}
// Program.cs:  BenchmarkRunner.Run<StringJoinBenchmarks>();

Run in Release: dotnet run -c Release.

Pitfalls the skill prevents

• Benchmarking in Debug, or without BenchmarkRunner (JIT/optimizations differ).
• Dead-code elimination — return the result so the work isn't optimized away.
• Setup work inside the [Benchmark] method — move it to [GlobalSetup].
• Comparing without a [Baseline] — you need a reference point.
• Tiny inputs that measure noise, not the algorithm.

Reading the summary

• Mean (ns/us/ms) — central tendency; check StdDev/Error for noise.
• Ratio — vs baseline (the headline comparison).
• Gen0/1/2 + Allocated — allocations and GC pressure; often the real story, not raw time.
• A faster method that allocates 10× more may lose under real load — weigh both.

Principles

• Measure the realistic input size and shape, not a toy.
• Allocations are a first-class result — keep [MemoryDiagnoser] on.
• One variable at a time; keep benchmarks isolated and deterministic.
• Don't micro-optimize a path that isn't hot — confirm it matters first (hotpath-profiler-assistant).