OKX Adapter Benchmarks

Numbers measured 2026-05-19 on AMD Ryzen Threadripper 9980X under rustc 1.95.0, bench-lto profile (release opts + lto = "fat" + codegen-units = 1, debug = full), ASLR disabled via setarch -R, default CPU governor.

Refresh on substantive perf change or before release; bump the date. Absolute numbers vary by machine; only same-machine deltas are meaningful.

How to reproduce

sudo cpupower frequency-set -g performance
setarch -R cargo bench -p nautilus-okx --profile bench-lto \
    --bench data --bench exec --bench micros --bench signing
sudo cpupower frequency-set -g powersave  # restore default

For policy and the general noise-reduction recipe see BENCHMARKING.md at the repo root.

Inbound pipeline (data.rs)

Raw WS frame bytes -> Nautilus domain type. Covers decode + parse + cache lookup + Nautilus type construction. No I/O, no async runtime, no channel.

Rows ordered from the most fundamental market-data stream (book deltas) down through derived streams (mark/index/funding/bars), then the private user streams (live order / fill) at the end.

Execution pipeline (exec.rs)

Strategy command (place/cancel/modify) -> wire bytes ready to send. Each iteration both constructs the request struct and serializes it to JSON, so the numbers reflect build + serialize together. OKX uses WebSocket for live order ops with no per-message signature (auth is established once at login); the per-request HMAC cost incurred by the HTTP path (instrument fetch, algo orders) is in signing.rs below.

submit_market, submit_limit, and submit_stop_market emit the HTTP order / order-algo request bodies (OKXPlaceOrderRequest, OKXPlaceAlgoOrderRequest). submit_ws_limit, cancel, and modify emit the production WS payload (OKXWsRequest<WsPostOrderParams>, WsCancelOrderParams, WsAmendOrderParams).

HTTP signing (signing.rs)

HMAC-SHA256 over (timestamp + method + path + body), base64-encoded. Only the HTTP path signs; the WS exec path does not.

Dispatch (exec.rs)

Venue execution report (FillReport, OrderStatusReport) forwarded via ExecutionEventEmitter. Measures the untracked report-fallback path through dispatch_execution_reports: trade-id dedup, dispatch-state bookkeeping, and send_*_report. The tracked-order path (dispatch_ws_message -> dispatch_parsed_order_event -> OrderAccepted/OrderFilled event construction) is pub(crate) and not exercised here; numbers below therefore exclude the per-event construction cost the tracked path adds.

Component breakdown (micros.rs)

Diagnostic benches that decompose the pipeline numbers above. Use these to localise where time goes when a pipeline bench regresses.

Notes

• Inbound is JSON-decode dominated. decode_only/book accounts for roughly 80% of inbound_pipeline/book_deltas (3.26 µs of 4.01 µs), and decode_only/trade accounts for roughly 90% of inbound_pipeline/trades (614 ns of 686 ns). Parsing itself is around 500 ns for a 10-level book delta and well under 100 ns for a trade tick.
• OKXWsFrame still buffers through serde_json::Value once before variant dispatch. Field extraction now takes ownership via Map::remove(...) instead of .cloned(), which removed the deep-clone per field. The remaining inbound headroom is the Value buffer itself; replacing it with a serde::de::Visitor or RawValue peek is the next lever and would also let OKXBookMsg levels deserialize without the intermediate Vec<Value> allocation.
• Exec is allocation-bound, not signature-bound. Build + JSON serialize lands in ~70-200 ns depending on shape, and OKX does not sign per-message on the WS path. HMAC-SHA256 (signing.rs) is HTTP-only.
• Dispatch runs against a fresh empty WsDispatchState each iteration. The state is built in the iter_batched setup closure (which Criterion excludes from timing), so the measured time is the dispatch_execution_reports body only. Production state lives forever and accumulates dedup entries; the steady-state cost on a reused state with a dedup hit is well below 100 ns.