Development documentation · Some APIs may be unreleased Published release →

Guides

Benchmarks

neoCLR benchmarks cover application request processing and general runtime execution. Each report identifies the measured build, workload, machine and execution mode.

Report What it measures
HTTP request processing Plain text and typed JSON across interpreter, JIT and AOT
Lazy JIT and application load Development startup, complex handlers, I/O pressure and sync/concurrent GC tradeoffs
Garbage collection Matched collector-path HTTP comparisons, allocation/reclamation diagnostics and correctness evidence
General execution CPU work, JSON serialization/deserialization and startup across interpreter, JIT and AOT

The retained Preview 18 macOS ARM64 and Windows x64 reports cover CPU/JSON execution and plain/typed-JSON HTTP. Each platform validates 55 CPU/JSON processes and 1,152 HTTP responses. These short workstation/hosted-runner series retain variability and limitations; they are not controlled-host regression gates or cross-platform rankings. Database and collector reports identify their own builds separately. Sustained cross-runtime measurements remain open.

Benchmark history preserves dated results and their raw evidence so future comparable runs can show progress, including regressions.

Reading the results

Startup and steady-state work are reported separately. Request reports include throughput, latency distributions and errors; execution reports state exactly what one operation includes. AOT build time and deployment size are separate from run time. Compare modes within the same recorded experiment, using equivalent inputs and validated outputs. Different revisions, machines or workloads are separate series.

Graphs will accompany qualified measurements, with units, labeled execution modes, visible variability and a numerical table. Each graph will link to raw samples, reproduction commands, toolchain revisions and environment details. Missing results will remain marked as pending, never plotted as zero. Comparisons with .NET, Go and JavaScript require equivalent applications and recorded configurations.

Measurement tools

The Raven benchmark framework and external runner support selectable cases and separate process startup and operation batches. Reports distinguish internal operation timing from external timings that include validation, protocol and scheduling costs.

Preview 19 five-round measurements and host conditions are retained in the release qualification record. The loaded macOS development host does not establish a controlled speedup.