Benchmarks

bench/SqlBound.Benchmarks compares SqlBound's generated code against Dapper and hand-written raw ADO.NET (the practical ceiling) on identical SQL over an in-memory SQLite database, with MemoryDiagnoser capturing allocations.

How to run

dotnet run --project bench/SqlBound.Benchmarks -c Release -- --filter "*"

Benchmarks are built — never run — by CI: numbers from shared runners are noise. The baseline below comes from a local run; re-run locally to compare like for like.

Reading the numbers

• Raw ADO.NET is the baseline (ratio 1.00) in every category: it is what a human writes by hand with GetOrdinal + typed getters, i.e. the same straight-line code the generator emits. SqlBound's goal is to match it, not beat it.
• Absolute numbers are dominated by SQLite itself; the meaningful signal is the relative time and allocation across the three approaches doing identical work.
• Allocations are where reflection-free codegen shows most clearly — Dapper's materializer is fast, but its dynamic parameter and mapping machinery allocates more.

Baseline results

Captured 2026-07-11 with the command above:

BenchmarkDotNet v0.15.8, Windows 11 (10.0.26200.8737/25H2)
AMD Ryzen Threadripper 3960X 3.80GHz, 1 CPU, 48 logical and 24 physical cores
.NET SDK 10.0.301, .NET 10.0.9, X64 RyuJIT x86-64-v3

Method	Categories	Mean	Ratio	Allocated	Alloc Ratio
RawAdoNet_Execute	Execute	3.349 us	1.00	960 B	1.00
SqlBound_Execute	Execute	3.652 us	1.09	1000 B	1.04
Dapper_Execute	Execute	4.234 us	1.26	1536 B	1.60
RawAdoNet_List	List1000	432.710 us	1.00	105832 B	1.00
SqlBound_List	List1000	517.365 us	1.20	105904 B	1.00
Dapper_List	List1000	592.187 us	1.37	151496 B	1.43
RawAdoNet_Scalar	Scalar	2.865 us	1.00	864 B	1.00
SqlBound_Scalar	Scalar	2.881 us	1.01	912 B	1.06
Dapper_Scalar	Scalar	2.987 us	1.04	864 B	1.00
RawAdoNet_Single	SingleRow	4.572 us	1.00	1624 B	1.00
SqlBound_Single	SingleRow	4.792 us	1.05	1696 B	1.04
Dapper_Single	SingleRow	5.607 us	1.23	2080 B	1.28

Takeaways from this baseline:

• SqlBound stays within 1–9% of hand-written ADO.NET on point operations and beats Dapper in every category on both time and allocations.
• List materialization allocates exactly what the hand-written loop allocates (1.00 ratio; Dapper allocates 1.43×) — the reflection-free straight-line codegen doing its job.
• The 1.20× list time (vs raw's 1.00) is the price of generated null guards: the generated code checks IsDBNull on every column to throw a descriptive error on unexpected NULLs, where the hand-written loop checks only the nullable one. That trade is deliberate.
• Two benchmark-fairness notes: the row type uses SQLite's natural provider types (long, double) because Dapper's constructor mapping requires exact type matches; and Dapper's strict QuerySingleAsync was chosen to match SqlBound's strict-single semantics.