Benchmarks

This page documents the active v0.3 engineering benchmark and the retained v0.2 raw-loop reference. Runtime evidence is host-dependent; deterministic receipts and tests remain the correctness claims.

v0.3 active-workload receipt

The tracked engineering_benchmark_v030.json measures both product reproduction and a materialized order/fill workload on Python 3.12.2, Apple arm64:

Benchmark	Result	What is exercised
Market Risk Case, median of 3 clean directories	23.7126 s	2,198-session report, bootstrap inference, SVGs, schemas, hashes
Active engine, 100,000 market events	38,767 events/s	10 symbols, target-weight resize, t+1 fills, commission ledger
Active engine peak traced memory	5.919 MB	Same 100,000-event workload

The three report runs produced the same receipt. This receipt is a conservative single-host reference, not a performance promise. The benchmark hashes its own source and the engine, execution, portfolio, and report sources so comparisons cannot silently change the workload.

PYTHONPATH=src python benchmarks/bench_v030.py \
  --output docs/assets/engineering_benchmark_v030.json

Retained v0.2 raw-loop reference

The earlier host-dependent receipt is tracked as benchmark.json:

Benchmark	Result	Environment
Audit Lab, median of 5 clean output directories	1.3745 s	Python 3.12.2, Apple arm64
Event loop, 1,000,000 no-op events	1,464,231 events/s	Python 3.12.2, Apple arm64

This no-op loop isolates dispatch overhead, but it does not exercise order sizing, future fill scheduling, or a cost ledger. It is retained as a narrow regression baseline rather than presented as application throughput.

• Script: benchmarks/bench_engine.py
• Purpose: Measures raw event throughput of the engine and Portfolio wiring under a no-op strategy and zero-cost execution model.

Running the raw loop locally

python benchmarks/bench_engine.py

The harness prints a small JSON with the number of processed events, wall-clock seconds, and events/sec.

Historical example on Apple M2 Pro (32GB, macOS 14.6.1):

{"events": 1000000, "sec": 0.773, "evps": 1294141}

Your numbers will vary by hardware, Python version, and build flags. Use the harness to compare relative changes across code revisions (e.g., after refactoring a tight loop).

Profiling a run

Enable profiling for any CLI run and inspect hotspots with snakeviz or gprof2dot:

microalpha run -c configs/meanrev.yaml --profile
# or
MICROALPHA_PROFILE=1 microalpha run -c configs/meanrev.yaml

The engine writes profile.pstats under the active run’s artifact directory:

artifacts/<run_id>/profile.pstats

Open it with snakeviz:

pip install snakeviz
snakeviz artifacts/<run_id>/profile.pstats

This integrates with the per-run artifacts so profiles travel alongside metrics and trades.