performance.md

Performance & High-Throughput

Benchmarks

Mode	Throughput	Latency
Sequential (cached)	250+ req/s	<100µs
Sequential (cold)	—	~3–4ms
Parallel / Rayon (8-core)	1,000+ req/s	—
Channel D semantic	—	~31µs

Node.js (async / concurrent)

napi-rs automatically runs evaluations on worker threads — Promise.all gives you free parallelism:

import { Firewall } from "policy-gate";

const firewall = await Firewall.create();

const results = await Promise.all(
  ["What is the capital of France?", "Write a function", "Hello!"]
    .map(p => firewall.evaluate(p))
);

Rust (parallel batch)

Enable the parallel feature for Rayon-based batch evaluation:

# Cargo.toml
firewall-core = { path = "./crates/firewall-core", features = ["parallel"] }

use firewall_core::evaluate_batch_parallel;

let results = evaluate_batch_parallel(vec![
    "What is the capital of France?".to_string(),
    "Write a function".to_string(),
], 0);

Running benchmarks

# All benchmark groups
cargo bench -p firewall-core

# Individual groups
cargo bench -p firewall-core --bench channel_a    # FSM + allowlist
cargo bench -p firewall-core --bench channel_b    # Rule engine
cargo bench -p firewall-core --bench voter        # 1oo2D decision logic
cargo bench -p firewall-core --bench integration  # Full pipeline
cargo bench -p firewall-core --bench normalise    # Unicode normalization

See crates/firewall-core/BENCHMARKS.md for detailed results.

BERT semantic mode (optional)

For deeper semantic analysis at the cost of 5–20ms latency:

pip install huggingface_hub
python scripts/setup_bert.py

Downloads required model files to models/. Enable in firewall.toml via engine_mode = "bert".