bench

Overview ¶

Concurrency / latency / write benchmark for api-log.

Spawns -conc workers; each fires -count requests cycling through the configured protocols. Records per-request latency and prints percentile and throughput stats; emits a JSON summary on stdout for the orchestrator to ingest.

Typical use:

bench -upstream http://localhost:7861 \
      -key sk-... -conc 50 -count 20 \
      -chat-model gpt-4o-mini \
      -anthropic-model claude-haiku-4-5 \
      -responses-model gpt-4o-mini

The bench is upstream-agnostic: point -upstream at the api-log proxy (default) to measure end-to-end including the proxy, or at the upstream directly to measure the bare baseline. Two runs with the same -seed give the same request body sequence, so before/after comparisons are apples-to-apples.