Skip to content

README.md

Latest commit

 

History

History
47 lines (31 loc) · 2.12 KB

README.md

File metadata and controls

47 lines (31 loc) · 2.12 KB
Raw

Verifier Micro-Benchmark

This benchmark measures the Rust verifier /verify endpoint only. It does not measure gateway, wallet UI, OpenRouter, Redis, or end-to-end x402 latency.

Requirements

  • wrk
  • bun
  • repo root dependencies installed with bun install
  • verifier running on http://127.0.0.1:3002 unless VERIFIER_URL is set

The script does not install wrk or any system package. Install tooling separately, then record the install method with your results.

Run

cd /path/to/MicroAI-Paygate
cd verifier && cargo run

In another shell:

RESULTS_FILE=bench/RESULTS-2026-05-13.txt bench/bench.sh

Useful overrides:

THREADS=4 CONNECTIONS=64 DURATION=60s EXPECTED_MAX_RPS=3000 bench/bench.sh
PAYLOAD_COUNT=200000 bench/bench.sh
VERIFIER_URL=http://127.0.0.1:3002 bench/bench.sh
SIGNATURE_EXPIRY_SECONDS=300 bench/bench.sh
BENCH_REVEAL_HOST=true bench/bench.sh

Method

bench.sh pre-generates one-time EIP-712 payment payloads with Bun and ethers, writes them to a temporary JSONL file, then gives wrk a Lua script that consumes each signed nonce at most once. This avoids benchmarking replay rejection instead of signature verification. By default, PAYLOAD_COUNT is derived from DURATION * EXPECTED_MAX_RPS; increase either value if the validity check reports exhausted payloads.

The script refuses runs where DURATION plus payload age would approach SIGNATURE_EXPIRY_SECONDS, and the Lua script aggregates per-thread counts for non-200 or is_valid:false responses. Set SIGNATURE_EXPIRY_SECONDS to the same value used by the verifier when benchmarking a non-default verifier configuration.

The default private key is a deterministic local test key used only for benchmark signing. Do not use a funded wallet.

Reporting

Only cite numbers that appear in a committed bench/RESULTS-*.txt file. Keep raw wrk output, non-identifying hardware metadata, thread/connection settings, duration, payload count, and validity-check counters together in the results file. Hostname and full uname -a output are omitted by default; set BENCH_REVEAL_HOST=true only for private local diagnostics.