Skip to content

BENCHMARKS.md

Latest commit

 

History

History
45 lines (35 loc) · 1.7 KB

BENCHMARKS.md

File metadata and controls

45 lines (35 loc) · 1.7 KB

Agent OS Performance Benchmarks

Benchmarked on: AMD Ryzen 9 5950X, 32GB RAM, Python 3.11

Summary

Operation Throughput Latency (p50) Latency (p99)
Kernel Execute (allow) 45,000 ops/sec 0.022ms 0.065ms
Kernel Execute (deny) 52,000 ops/sec 0.019ms 0.058ms
Policy Evaluation (10 rules) 120,000 ops/sec 0.008ms 0.025ms
Policy Evaluation (100 rules) 18,000 ops/sec 0.055ms 0.15ms
YAML Policy Load 2,800 ops/sec 0.36ms 0.95ms
Audit Entry Write 145,000 ops/sec 0.007ms 0.021ms
Audit Log Query (10K entries) 1,200 ops/sec 0.83ms 2.1ms
Circuit Breaker Check 890,000 ops/sec 0.001ms 0.003ms
Adapter Overhead (avg) — 0.015ms 0.042ms
Full Governed Action 28,000 ops/sec 0.036ms 0.098ms

Key Takeaways

  • Sub-100μs governance: Full kernel enforcement in <0.1ms p99
  • Zero-overhead adapters: Framework integration adds <42μs
  • 1,577+ tests passing with governance enforcement active

Running Benchmarks

# Run all benchmarks with JSON + markdown output
python -m benchmarks.run_all

# Run individual benchmark modules
python -m benchmarks.bench_kernel
python -m benchmarks.bench_policy
python -m benchmarks.bench_audit
python -m benchmarks.bench_adapters

Methodology

  • Each operation is measured over thousands of iterations
  • Latencies captured via time.perf_counter() (sub-microsecond resolution)
  • p50/p95/p99 percentiles computed from sorted latency distributions
  • Concurrent benchmarks use asyncio.gather() to simulate real-world load
  • All benchmarks run against in-memory backends to isolate kernel overhead