Skip to content

RESULTS.md

Latest commit

 

History

History
51 lines (34 loc) · 2.49 KB

RESULTS.md

File metadata and controls

51 lines (34 loc) · 2.49 KB

Benchmark Results: Lua vs Go FFI OpenAPI Validator

Environment

  • CPU: AMD EPYC (host machine)
  • OpenResty: 1.21.4.4
  • Worker processes: 1
  • Tool: wrk2 (4 threads, 50 connections, 10s duration, target 50K req/s)

Test Scenario

  • Endpoint: POST /users/{userId}?limit=10
  • Path param: userId (integer, minimum 1)
  • Query param: limit (integer, 1-100)
  • Header: X-Request-Id (required, string)
  • Body: ~300B JSON with 10 fields (string, integer, number, boolean, array, nested object, additionalProperties, enum, pattern)
  • Spec: OpenAPI 3.0 with comprehensive schema constraints

Results

Metric Lua Validator Go FFI Validator Improvement
Throughput 26,993 req/s 18,610 req/s +45%
p50 Latency 2.33s 3.14s -35%
p99 Latency 4.53s 6.19s -27%
Max Latency 4.59s 6.26s -27%

Note: High absolute latencies are due to wrk2's coordinated omission correction at a 50K req/s target rate that exceeds the server's max throughput on a single worker. The relative comparison is what matters.

Analysis

The Lua validator achieves ~45% higher throughput than the Go FFI version on the same spec and request payload. Key reasons:

  1. No FFI boundary crossing: The Go FFI version serializes headers as JSON, copies body as C string, crosses the cgo boundary, then deserializes everything in Go. The Lua version operates on native Lua tables directly.

  2. No Go GC interference: The Go runtime's garbage collector runs in the same process as OpenResty workers. Under load, GC pauses add latency variance. The Lua version runs entirely under LuaJIT's lightweight allocator.

  3. Pre-compiled validation: The Lua version pre-compiles jsonschema validators during compile(). Each request only runs the pre-built validation function, with zero reflection or dynamic dispatch.

  4. Zero-copy parameters: Path, query, and header parameters are read directly from OpenResty's request API. The Go FFI version re-parses the full URL and re-creates an http.Request object on every call.

Single-thread Microbenchmark

Running 50K iterations in a tight loop (os.clock CPU time):

Metric Lua Validator Go FFI Validator
Throughput ~120K ops/s ~110K ops/s
Avg latency 0.008 ms/op 0.009 ms/op

The single-thread difference is smaller (~1.1-1.2x) because both versions avoid the concurrency-related overheads. The gap widens under concurrent load.