Skip to content
Purview Dev logo

Documentation section: Reference

Performance

Stable Telemetry SourceGenerator Reviewed 2026-11-15 purview-dev/telemetry-sourcegenerator Star on GitHub 37 activity c-sharp distributed-tracing dotnet events high-performance-logging instrumentation logging melt metrics nuget observability open-telemetry open-telemetry-csharp otel roslyn source-generator spans telemetry tracing

Full cross-runtime benchmark results for the Purview Telemetry Source Generator, generated by BenchmarkDotNet from the benchmarks/ project.

Terminal window
dotnet run --project benchmarks/Purview.Telemetry.Benchmarks/Purview.Telemetry.Benchmarks.csproj \
--configuration Release --framework net11.0

just benchmark runs the same command. Every benchmark class declares one job per runtime (RuntimeMoniker.Net10_0, Net11_0, Net48), so one invocation measures all three and the framework the suite is launched with only selects the BenchmarkDotNet host.

Results are written to BenchmarkDotNet.Artifacts/results/ as *-report-github.md, *.csv, and *.html. That folder is gitignored: the artifacts stay on the machine that produced them, and the numbers below are what is published.

BenchmarkDotNet v0.16.0-preview.2, Windows 11 (10.0.26300.9974)
13th Gen Intel Core i9-13900KF 3.00GHz, 1 CPU, 32 logical and 24 physical cores
Memory: 63.77 GB Total, 39.06 GB Available
.NET SDK 11.0.100
[Host] : .NET 11.0.0 (11.0.0, 11.0.26.53018), X64 RyuJIT x86-64-v3
Job-GVKUBM : .NET 10.0.15 (10.0.15, 10.0.1526.53008), X64 RyuJIT x86-64-v3
Job-IHFIKV : .NET 11.0.0 (11.0.0, 11.0.26.53018), X64 RyuJIT x86-64-v3
Job-UGENHH : .NET Framework 4.8 (4.8.10008.0), X64 RyuJIT VectorSize=256

Every table below comes from this machine and SDK. ActivityBenchmarks, MultiTargetVsSingleTargetBenchmarks, LoggerMultiTargetBenchmarks, TagListBenchmarks and GeneratorModelBenchmarks come from the full suite. MetricsBenchmarks and LoggerBenchmarks were re-run afterwards, once their hand-written baselines were changed to implement the benchmark interface so both sides of those comparisons are consumed through the same call shape; before that they compared an interface-typed generated instance against a concrete-typed hand-written one, which reports interface dispatch as generator overhead. A table never mixes runs. Ratio is measured against the row marked in bold, which is the class’s Baseline = true row. Allocated - means the path does not allocate. Rows whose mean is a nanosecond or two carry run-to-run variation of a few tenths of a nanosecond; treat ratios in that range as “indistinguishable” rather than as a ranking.

Generated hot-path methods carry [MethodImpl(MethodImplOptions.AggressiveInlining)], which lets the JIT inline the small generated telemetry methods into callers.

Source: ActivityBenchmarks

Compares the source-generator-produced ActivityOnlyTelemetryCore against a hand-written ManualActivityTelemetry under two conditions: no listener registered (fast-path) and a full-sampling ActivityListener active (production path).

Method Runtime HasListener Mean Ratio Allocated
Manual: start + complete .NET 10.0 False 1.504 ns 1.00 -
Generated: start + complete .NET 10.0 False 1.514 ns 1.01 -
Manual: start + fail .NET 10.0 False 1.355 ns 0.90 -
Generated: start + fail .NET 10.0 False 1.730 ns 1.15 -
Manual: start + complete .NET 11.0 False 1.340 ns 1.00 -
Generated: start + complete .NET 11.0 False 1.712 ns 1.28 -
Manual: start + fail .NET 11.0 False 1.531 ns 1.14 -
Generated: start + fail .NET 11.0 False 1.544 ns 1.15 -
Manual: start + complete .NET Framework 4.8 False 16.646 ns 1.00 -
Generated: start + complete .NET Framework 4.8 False 16.836 ns 1.01 -
Manual: start + fail .NET Framework 4.8 False 16.117 ns 0.97 -
Generated: start + fail .NET Framework 4.8 False 16.768 ns 1.01 -
Manual: start + complete .NET 10.0 True 218.245 ns 1.00 1008 B
Generated: start + complete .NET 10.0 True 212.817 ns 0.98 1008 B
Manual: start + fail .NET 10.0 True 192.314 ns 0.88 920 B
Generated: start + fail .NET 10.0 True 191.271 ns 0.88 920 B
Manual: start + complete .NET 11.0 True 204.501 ns 1.00 1008 B
Generated: start + complete .NET 11.0 True 195.388 ns 0.96 1008 B
Manual: start + fail .NET 11.0 True 201.220 ns 0.98 920 B
Generated: start + fail .NET 11.0 True 193.368 ns 0.95 920 B
Manual: start + complete .NET Framework 4.8 True 524.705 ns 1.00 1268 B
Generated: start + complete .NET Framework 4.8 True 520.174 ns 0.99 1268 B
Manual: start + fail .NET Framework 4.8 True 500.322 ns 0.95 1179 B
Generated: start + fail .NET Framework 4.8 True 498.376 ns 0.95 1180 B

Interpretation: Generated activities are indistinguishable from hand-written ones on every runtime. With a listener active — the production path — the generated rows land within 1–5% of the hand-written rows (0.95–0.99), and the absolute time (190–525 ns) is dominated by the framework’s own Activity creation and event recording, not by generated code. The one-byte difference in the .NET Framework start + fail row (1179 B vs 1180 B) is allocation-size rounding of the same event payload. With no listener both implementations are a single guard: ~1.3–1.7 ns on .NET 10/11, where a few tenths of a nanosecond is run-to-run noise, and ~16–17 ns on .NET Framework. Nothing allocates on the no-listener path on any runtime.

Source: LoggerBenchmarks

Compares three logging approaches: hand-written LoggerMessage.Define (gold-standard manual), generated v1 (LoggerMessage.Define pattern), and generated v2 (state-based ThreadLocalState pattern). HasLogging = true uses an always-enabled no-op logger, so the full path — IsEnabled check, state population, ILogger.Log dispatch — is exercised; false uses NullLogger, so the guard short-circuits.

Method Runtime HasLogging Mean Ratio Allocated
Manual (LoggerMessage.Define) — single Info call .NET 10.0 False 1.133 ns 1.00 -
Generated v1 (LoggerMessage.Define) — single Info call .NET 10.0 False 1.106 ns 0.98 -
Generated v2 (ThreadLocalState) — single Info call .NET 10.0 False 1.114 ns 0.98 -
Manual (LoggerMessage.Define) — full lifecycle (4 calls) .NET 10.0 False 1.492 ns 1.32 -
Generated v1 (LoggerMessage.Define) — full lifecycle (4 calls) .NET 10.0 False 1.500 ns 1.32 -
Generated v2 (ThreadLocalState) — full lifecycle (4 calls) .NET 10.0 False 1.374 ns 1.21 -
Manual (LoggerMessage.Define) — single Info call .NET 11.0 False 1.116 ns 1.00 -
Generated v1 (LoggerMessage.Define) — single Info call .NET 11.0 False 1.129 ns 1.01 -
Generated v2 (ThreadLocalState) — single Info call .NET 11.0 False 1.116 ns 1.00 -
Manual (LoggerMessage.Define) — full lifecycle (4 calls) .NET 11.0 False 1.532 ns 1.37 -
Generated v1 (LoggerMessage.Define) — full lifecycle (4 calls) .NET 11.0 False 1.494 ns 1.34 -
Generated v2 (ThreadLocalState) — full lifecycle (4 calls) .NET 11.0 False 1.398 ns 1.25 -
Manual (LoggerMessage.Define) — single Info call .NET Framework 4.8 False 2.819 ns 1.00 -
Generated v1 (LoggerMessage.Define) — single Info call .NET Framework 4.8 False 2.816 ns 1.00 -
Generated v2 (ThreadLocalState) — single Info call .NET Framework 4.8 False 8.189 ns 2.91 -
Manual (LoggerMessage.Define) — full lifecycle (4 calls) .NET Framework 4.8 False 11.380 ns 4.04 -
Generated v1 (LoggerMessage.Define) — full lifecycle (4 calls) .NET Framework 4.8 False 11.396 ns 4.04 -
Generated v2 (ThreadLocalState) — full lifecycle (4 calls) .NET Framework 4.8 False 18.976 ns 6.73 -
Manual (LoggerMessage.Define) — single Info call .NET 10.0 True 5.355 ns 1.00 -
Generated v1 (LoggerMessage.Define) — single Info call .NET 10.0 True 5.360 ns 1.00 -
Generated v2 (ThreadLocalState) — single Info call .NET 10.0 True 4.391 ns 0.82 -
Manual (LoggerMessage.Define) — full lifecycle (4 calls) .NET 10.0 True 20.208 ns 3.77 -
Generated v1 (LoggerMessage.Define) — full lifecycle (4 calls) .NET 10.0 True 19.600 ns 3.66 -
Generated v2 (ThreadLocalState) — full lifecycle (4 calls) .NET 10.0 True 17.964 ns 3.35 -
Manual (LoggerMessage.Define) — single Info call .NET 11.0 True 5.352 ns 1.00 -
Generated v1 (LoggerMessage.Define) — single Info call .NET 11.0 True 5.315 ns 0.99 -
Generated v2 (ThreadLocalState) — single Info call .NET 11.0 True 4.416 ns 0.83 -
Manual (LoggerMessage.Define) — full lifecycle (4 calls) .NET 11.0 True 18.554 ns 3.47 -
Generated v1 (LoggerMessage.Define) — full lifecycle (4 calls) .NET 11.0 True 19.103 ns 3.57 -
Generated v2 (ThreadLocalState) — full lifecycle (4 calls) .NET 11.0 True 15.708 ns 2.94 -
Manual (LoggerMessage.Define) — single Info call .NET Framework 4.8 True 17.785 ns 1.00 -
Generated v1 (LoggerMessage.Define) — single Info call .NET Framework 4.8 True 18.235 ns 1.03 -
Generated v2 (ThreadLocalState) — single Info call .NET Framework 4.8 True 14.652 ns 0.82 -
Manual (LoggerMessage.Define) — full lifecycle (4 calls) .NET Framework 4.8 True 71.259 ns 4.01 -
Generated v1 (LoggerMessage.Define) — full lifecycle (4 calls) .NET Framework 4.8 True 71.375 ns 4.01 -
Generated v2 (ThreadLocalState) — full lifecycle (4 calls) .NET Framework 4.8 True 45.316 ns 2.55 -

Interpretation: Every row allocates zero bytes, on every runtime and on both paths. All three implementations are consumed through ILoggerOnlyTelemetry, so neither side is inlined into the measurement loop. On the full path (HasLogging = true) generated v2 is the fastest implementation on all three runtimes — 4.39 ns against 5.36 ns hand-written on .NET 10.0 (0.82), 4.42 against 5.35 on .NET 11.0 (0.83), 14.65 against 17.79 on .NET Framework 4.8 (0.82) — while generated v1, which reproduces the hand-written LoggerMessage.Define delegate, is at parity (0.99–1.03). The four-call lifecycle costs v2 15.7–18.0 ns against 18.6–20.2 ns hand-written on the modern runtimes, and v1 sits within 3% of the hand-written lifecycle. On the disabled path the IsEnabled guard dominates and the two delegate-based implementations are indistinguishable (1.11–1.13 ns single call on .NET 10/11, 2.82 ns on .NET Framework). The one row where v2’s wiring is measurably more expensive is that disabled single call on .NET Framework 4.8: 8.19 ns against 2.82 ns for the delegate baseline, 2.91× before any log state is built.

Source: MultiTargetVsSingleTargetBenchmarks

Measures the overhead of emitting Activity + Logging + Metrics from a single method call (multi-target) vs. Activity-only (single-target), comparing generated and manual code.

Method Runtime HasListener Mean Ratio Allocated Alloc Ratio
Single-target (generated): start + complete .NET 10.0 False 1.716 ns 1.00 - NA
Multi-target (generated): start + complete .NET 10.0 False 11.346 ns 6.61 24 B NA
Multi-target (manual): start + complete .NET 10.0 False 14.384 ns 8.38 24 B NA
Multi-target (generated): start + complete + record latency .NET 10.0 False 11.898 ns 6.94 24 B NA
Multi-target (manual): start + complete + record latency .NET 10.0 False 13.162 ns 7.67 24 B NA
Single-target (generated): start + complete .NET 11.0 False 1.693 ns 1.00 - NA
Multi-target (generated): start + complete .NET 11.0 False 11.704 ns 6.92 24 B NA
Multi-target (manual): start + complete .NET 11.0 False 12.942 ns 7.65 24 B NA
Multi-target (generated): start + complete + record latency .NET 11.0 False 10.806 ns 6.39 24 B NA
Multi-target (manual): start + complete + record latency .NET 11.0 False 13.327 ns 7.88 24 B NA
Single-target (generated): start + complete .NET Framework 4.8 False 16.534 ns 1.00 - NA
Multi-target (generated): start + complete .NET Framework 4.8 False 63.225 ns 3.82 24 B NA
Multi-target (manual): start + complete .NET Framework 4.8 False 73.849 ns 4.47 24 B NA
Multi-target (generated): start + complete + record latency .NET Framework 4.8 False 81.570 ns 4.93 24 B NA
Multi-target (manual): start + complete + record latency .NET Framework 4.8 False 88.946 ns 5.38 24 B NA
Single-target (generated): start + complete .NET 10.0 True 204.268 ns 1.00 1008 B 1.00
Multi-target (generated): start + complete .NET 10.0 True 212.584 ns 1.04 1032 B 1.02
Multi-target (manual): start + complete .NET 10.0 True 236.564 ns 1.16 1032 B 1.02
Multi-target (generated): start + complete + record latency .NET 10.0 True 213.345 ns 1.04 1032 B 1.02
Multi-target (manual): start + complete + record latency .NET 10.0 True 216.492 ns 1.06 1032 B 1.02
Single-target (generated): start + complete .NET 11.0 True 209.172 ns 1.00 1008 B 1.00
Multi-target (generated): start + complete .NET 11.0 True 241.172 ns 1.15 1032 B 1.02
Multi-target (manual): start + complete .NET 11.0 True 204.857 ns 0.98 1032 B 1.02
Multi-target (generated): start + complete + record latency .NET 11.0 True 231.273 ns 1.11 1032 B 1.02
Multi-target (manual): start + complete + record latency .NET 11.0 True 232.157 ns 1.11 1032 B 1.02
Single-target (generated): start + complete .NET Framework 4.8 True 523.480 ns 1.00 1268 B 1.00
Multi-target (generated): start + complete .NET Framework 4.8 True 594.085 ns 1.14 1292 B 1.02
Multi-target (manual): start + complete .NET Framework 4.8 True 590.657 ns 1.13 1284 B 1.01
Multi-target (generated): start + complete + record latency .NET Framework 4.8 True 595.731 ns 1.14 1284 B 1.01
Multi-target (manual): start + complete + record latency .NET Framework 4.8 True 604.441 ns 1.16 1292 B 1.02

Interpretation: With a listener active — the production path — multi-target generation adds 4% over Activity-only on .NET 10.0 (1.04) and 11–15% on .NET 11.0 and .NET Framework 4.8, for the log call and metric increment it performs. On .NET 10.0 the generated multi-target code is also faster than the hand-written multi-target baseline (212.6 vs 236.6 ns), and on .NET Framework 4.8 the two are at parity (594.1 vs 590.7 ns). The hand-written multi-target baseline in this family is still called on its concrete type, which favours it, so where the generated row wins it wins despite that. On .NET 11.0 the generated row carries a ±18 ns standard deviation in this run while the hand-written row carries ±2 ns, so the apparent gap there is not resolvable. Without a listener the multi-target path costs 10.8–11.9 ns on the modern runtimes against 1.7 ns for the activity-only guard, because a log guard and a metric increment run even when no activity is sampled; the hand-written equivalent is slower still (12.9–14.4 ns). The 24 B per call on that path is the same on both implementations, and the listener-active rows allocate 1032 B against 1008 B for activity-only — the extra 24 B belongs to the metric and log state, not to the generated wrapper. See the Generated Output page for what the multi-target implementation looks like.

Source: LoggerMultiTargetBenchmarks

Compares single-target logging-only against multi-target (Activity + Logging + Metrics) generated code with a hand-written multi-target baseline. Logging is always enabled; HasListener controls activity sampling.

Method Runtime HasListener Mean Ratio Allocated Alloc Ratio
Multi-target (manual): start + complete .NET 10.0 False 12.91 ns 1.00 24 B 1.00
Multi-target (generated v1): start + complete .NET 10.0 False 13.18 ns 1.02 24 B 1.00
Multi-target (generated v2): start + complete .NET 10.0 False 11.10 ns 0.86 24 B 1.00
Multi-target (manual): full lifecycle .NET 10.0 False 13.60 ns 1.05 24 B 1.00
Multi-target (generated v1): full lifecycle .NET 10.0 False 13.53 ns 1.05 24 B 1.00
Multi-target (generated v2): full lifecycle .NET 10.0 False 12.84 ns 0.99 24 B 1.00
Single-target (generated v1): full lifecycle .NET 10.0 False 18.14 ns 1.40 - 0.00
Single-target (generated v2): full lifecycle .NET 10.0 False 16.63 ns 1.29 - 0.00
Multi-target (manual): start + complete .NET 11.0 False 13.33 ns 1.00 24 B 1.00
Multi-target (generated v1): start + complete .NET 11.0 False 13.15 ns 0.99 24 B 1.00
Multi-target (generated v2): start + complete .NET 11.0 False 10.54 ns 0.79 24 B 1.00
Multi-target (manual): full lifecycle .NET 11.0 False 13.24 ns 0.99 24 B 1.00
Multi-target (generated v1): full lifecycle .NET 11.0 False 13.78 ns 1.03 24 B 1.00
Multi-target (generated v2): full lifecycle .NET 11.0 False 11.92 ns 0.89 24 B 1.00
Single-target (generated v1): full lifecycle .NET 11.0 False 18.61 ns 1.40 - 0.00
Single-target (generated v2): full lifecycle .NET 11.0 False 15.43 ns 1.16 - 0.00
Multi-target (manual): start + complete .NET Framework 4.8 False 70.44 ns 1.00 24 B 1.00
Multi-target (generated v1): start + complete .NET Framework 4.8 False 72.88 ns 1.03 24 B 1.00
Multi-target (generated v2): start + complete .NET Framework 4.8 False 60.74 ns 0.86 24 B 1.00
Multi-target (manual): full lifecycle .NET Framework 4.8 False 88.92 ns 1.26 24 B 1.00
Multi-target (generated v1): full lifecycle .NET Framework 4.8 False 92.15 ns 1.31 24 B 1.00
Multi-target (generated v2): full lifecycle .NET Framework 4.8 False 80.08 ns 1.14 24 B 1.00
Single-target (generated v1): full lifecycle .NET Framework 4.8 False 73.28 ns 1.04 - 0.00
Single-target (generated v2): full lifecycle .NET Framework 4.8 False 45.30 ns 0.64 - 0.00
Multi-target (manual): start + complete .NET 10.0 True 238.98 ns 1.00 1032 B 1.00
Multi-target (generated v1): start + complete .NET 10.0 True 246.79 ns 1.03 1032 B 1.00
Multi-target (generated v2): start + complete .NET 10.0 True 220.70 ns 0.92 1032 B 1.00
Multi-target (manual): full lifecycle .NET 10.0 True 222.75 ns 0.93 1032 B 1.00
Multi-target (generated v1): full lifecycle .NET 10.0 True 219.33 ns 0.92 1032 B 1.00
Multi-target (generated v2): full lifecycle .NET 10.0 True 216.81 ns 0.91 1032 B 1.00
Single-target (generated v1): full lifecycle .NET 10.0 True 18.75 ns 0.08 - 0.00
Single-target (generated v2): full lifecycle .NET 10.0 True 15.28 ns 0.06 - 0.00
Multi-target (manual): start + complete .NET 11.0 True 223.96 ns 1.00 1032 B 1.00
Multi-target (generated v1): start + complete .NET 11.0 True 222.29 ns 0.99 1032 B 1.00
Multi-target (generated v2): start + complete .NET 11.0 True 212.34 ns 0.95 1032 B 1.00
Multi-target (manual): full lifecycle .NET 11.0 True 239.89 ns 1.07 1032 B 1.00
Multi-target (generated v1): full lifecycle .NET 11.0 True 233.14 ns 1.04 1032 B 1.00
Multi-target (generated v2): full lifecycle .NET 11.0 True 214.18 ns 0.96 1032 B 1.00
Single-target (generated v1): full lifecycle .NET 11.0 True 20.72 ns 0.09 - 0.00
Single-target (generated v2): full lifecycle .NET 11.0 True 16.17 ns 0.07 - 0.00
Multi-target (manual): start + complete .NET Framework 4.8 True 592.06 ns 1.00 1292 B 1.00
Multi-target (generated v1): start + complete .NET Framework 4.8 True 579.08 ns 0.98 1292 B 1.00
Multi-target (generated v2): start + complete .NET Framework 4.8 True 561.83 ns 0.95 1292 B 1.00
Multi-target (manual): full lifecycle .NET Framework 4.8 True 592.02 ns 1.00 1292 B 1.00
Multi-target (generated v1): full lifecycle .NET Framework 4.8 True 596.52 ns 1.01 1292 B 1.00
Multi-target (generated v2): full lifecycle .NET Framework 4.8 True 577.82 ns 0.98 1292 B 1.00
Single-target (generated v1): full lifecycle .NET Framework 4.8 True 72.80 ns 0.12 - 0.00
Single-target (generated v2): full lifecycle .NET Framework 4.8 True 45.37 ns 0.08 - 0.00

Interpretation: Generated and hand-written multi-target implementations allocate identically (1032 B with a listener active, 24 B without) and land within 9% of each other per scenario on every runtime; v2 is the fastest of the six multi-target rows on each of the three runtimes (0.79–0.95 against the manual start-complete baseline). Logging-only accounts for 15–21 ns of the multi-target cost (the two single-target rows: 15.3–20.7 ns on the modern runtimes, 45.4–72.8 ns on .NET Framework), so with a listener active the ~220–600 ns total is Activity creation, not logging or metrics. Without a listener the multi-target path still builds the metric and log state — 10.5–13.8 ns on .NET 10/11, less than the 15.4–18.6 ns the four logging-only calls cost, because a multi-target method emits fewer log events.

Source: MetricsBenchmarks and TagListBenchmarks

Compares the generated synchronous instruments against hand-written best-practice code, one instrument call per benchmark. All instruments are 0 allocations on every runtime, on both the generated and the hand-written path.

Scenario Manual (.NET 10.0) Generated (.NET 10.0) Manual (.NET 11.0) Generated (.NET 11.0) Manual (.NET Framework 4.8) Generated (.NET Framework 4.8)
auto-counter (0 tags) 1.29 ns 1.31 ns 1.33 ns 1.30 ns 16.16 ns 16.32 ns
auto-counter (1 tag) 1.32 ns 1.30 ns 1.28 ns 1.29 ns 19.75 ns 20.31 ns
up-down counter 1.34 ns 1.30 ns 1.28 ns 1.28 ns 17.03 ns 16.86 ns
histogram (0 tags) 1.28 ns 1.29 ns 1.29 ns 1.29 ns 16.97 ns 16.67 ns
histogram (1 tag) 1.28 ns 1.29 ns 1.31 ns 1.29 ns 19.58 ns 19.68 ns

Both implementations are consumed through IMetricsBenchmarkTelemetry, so neither side is inlined into the measurement loop and the comparison is like for like. On .NET 10.0 and .NET 11.0 the generated rows land within 2% of the hand-written ones — 1.29–1.31 ns against 1.28–1.34 ns — and on .NET Framework 4.8 within 3% (16.3–20.3 ns against 16.2–19.8 ns). Both sides sit at the same ~1.3 ns per update on the modern runtimes whatever the tag count: at that scale the instrument call is the cost, and the generator adds nothing measurable to it.

The source generator uses a tag-count optimization: methods with fewer than 4 tags pass inline KeyValuePair parameters (no heap allocation), while methods with 4 or more tags use a stack-allocated TagList struct.

Scenario .NET 10.0 .NET 11.0 .NET Framework 4.8
0 tags: histogram record 1.42 ns 1.35 ns 16.95 ns
1 tag: auto-counter add 1.37 ns 1.33 ns 20.06 ns
3 tags: histogram record 1.94 ns 1.85 ns 30.41 ns
4 tags (TagList): auto-counter add 5.28 ns 4.93 ns 48.32 ns
5 tags (TagList): auto-counter add 6.54 ns 5.24 ns 57.15 ns
6 tags (TagList): histogram record 7.10 ns 5.34 ns 60.56 ns

The TagList path measures 4.9–7.1 ns against 1.3–1.9 ns for the inline paths on the modern runtimes, and 48–61 ns against 17–30 ns on .NET Framework, but both remain allocation-free and in the single-digit-nanosecond range (tens of nanoseconds on .NET Framework). The cost tracks the number of tags on every runtime, which is the struct population itself rather than any allocation.

Source: GeneratorModelBenchmarks

Compares the two modelling strategies the generator’s pipeline can use for its tagged hierarchies (closed record InstrumentAttributeRecord / ActivityMethodTarget): an abstract record with one sealed case per instrument and virtual properties, versus a readonly struct union carrying a tag plus the case payloads with a non-boxing TryGetValue accessor. Both build the same number of models and read the same four facts off each one, which is what a generation pass does per method. This is generator-time work, not a consumer hot path: it is included so the decision recorded in Platform Support rests on a measurement.

Method Runtime MethodCount Mean Ratio Gen0 Allocated Alloc Ratio
closed record hierarchy: build + dispatch .NET 10.0 8 80.03 ns 1.00 0.0267 504 B 1.00
value-based struct union: build + dispatch .NET 10.0 8 41.25 ns 0.52 0.0200 376 B 0.75
closed record hierarchy: build + dispatch .NET 11.0 8 76.45 ns 1.00 0.0267 504 B 1.00
value-based struct union: build + dispatch .NET 11.0 8 41.58 ns 0.54 0.0200 376 B 0.75
closed record hierarchy: build + dispatch .NET Framework 4.8 8 130.01 ns 1.00 0.0815 514 B 1.00
value-based struct union: build + dispatch .NET Framework 4.8 8 193.11 ns 1.49 0.0610 385 B 0.75
closed record hierarchy: build + dispatch .NET 10.0 64 631.35 ns 1.00 0.1926 3640 B 1.00
value-based struct union: build + dispatch .NET 10.0 64 303.11 ns 0.48 0.1388 2616 B 0.72
closed record hierarchy: build + dispatch .NET 11.0 64 530.27 ns 1.00 0.1926 3640 B 1.00
value-based struct union: build + dispatch .NET 11.0 64 295.96 ns 0.56 0.1388 2616 B 0.72
closed record hierarchy: build + dispatch .NET Framework 4.8 64 967.34 ns 1.00 0.5798 3659 B 1.00
value-based struct union: build + dispatch .NET Framework 4.8 64 1,545.25 ns 1.60 0.4177 2633 B 0.72

Interpretation: On .NET 10 and .NET 11 the value-based union is roughly twice as fast (0.48–0.56) and allocates 25–28% less. On .NET Framework 4.8 it is the slower of the two, 1.49× at 8 methods and 1.60× at 64, because the struct’s copy and tag-switch cost exceeds what the record hierarchy pays there. That is the direction that decided the question: the records stayed, because the win on the modern runtimes is 35–39 ns at 8 methods and 234–328 ns at 64 methods per generation pass — a few nanoseconds per method, against a pass that also resolves every symbol, builds the message templates and writes the source text — and the change would make .NET Framework generation measurably worse. The exhaustive switch over a closed hierarchy, checked by the compiler, is the second reason.

ObservableCounter, ObservableGauge, and ObservableUpDownCounter are not benchmarked because they have no per-operation hot path to compare. They register a callback once at construction time (via meter.CreateObservable*) and are polled by the metrics collection pipeline — the generator produces the identical CreateObservable* call you would write by hand, so there is no wrapper layer on the measurement path.

The BenchmarkDotNet.Artifacts/ folder is gitignored, so the CSV and HTML artifacts are not published — they live only on the machine that ran the suite. To produce them, run the command in Reproducing; one invocation writes a -report-github.md, a -report.csv and a -report.html per benchmark class, and the CSVs carry the Error, StdDev and Median columns the tables above omit.