Measure first: the benchmark discipline

Measure first: the benchmark discipline

The commandment: no optimization without a measured baseline. The discipline that makes the measurement mean something:

// The minimal honest harness (what the checkpoint grades):
var work = (Func<int>)(() => SumTo(1000));
_ = work();                                     // 1. WARMUP — JIT, caches
var before = GC.GetAllocatedBytesForCurrentThread();   // 2. BASELINE
var sw = Stopwatch.StartNew();                  // 3. TIMED WINDOW
for (int i = 0; i < N; i++) _ = work();
sw.Stop();
var allocDelta = GC.GetAllocatedBytesForCurrentThread() - before;
Console.WriteLine($"{sw.Elapsed.TotalMicroseconds / N} us/op, {allocDelta / N} B/op");

The rules and the lies each prevents:

| Rule | Prevents | |---|---| | Warmup calls | First-call JIT + cache costs dominating N=1 runs | | Many iterations (100–10k) | Timer resolution noise (Stopwatch ≈ tens of ns, but scheduling is µs) | | Allocation delta, not wall time alone | CPU frequency scaling, machine noise faking speedups | | Repeats with min/median | One unlucky scheduling blob poisoning the run | | Same machine state (no debug build!) | Comparing debug vs release, or throttled laptop vs plugged |

Production-grade benchmarking uses BenchmarkDotNet, which isolates each benchmark in its own process, does pilot-phase iteration sizing, and reports statistical summaries — but its PHILOSOPHY is the table above, and the sandbox (no NuGet restore) can run the minimal harness instead.

The equal-opportunity failure: measuring the WRONG THING. Optimizing parsing when the request spends 90% of its time in DB round-trips is the classic. Profile the END-TO-END path before micro-optimizing any component.