Control UI protocol load

Build with pnpm build. The Linux benchmark creates private Gateway state, synthetic signed devices/credentials, loopback listeners, and a mock provider. Supply a delegated cgroup-v2 parent with CPU accounting; run the controller inside that subtree so Gateway migration is permitted. The benchmark owns a unique child cgroup and verifies cleanup. Use an isolated runner, never a live service cgroup.

taskset --cpu-list 4-31 node scripts/bench-gateway-concurrency.ts \  --control-ui-clients 50 --active-clients 25 --drivers 4 \  --duration-ms 60000 --timeout-ms 180000 --runs 1 --warmup 0 \  --gateway-runtime /path/to/node --gateway-cpus 0,1,2,3 \  --resource-cgroup /sys/fs/cgroup/benchmark-parent \  --transpiler-cache /benchmark-cache/node \  --output /benchmark-results/node-50-25.json

For 100/50, change the two client counts. Keep the Node driver unchanged when selecting another Gateway runtime. Record executable hashes, source/build/lockfile identities, host topology, kernel, and observed affinity with each campaign.

Workload contract

Drivers partition clients round-robin. Each signed Control UI webchat connection subscribes to inventory and its own session, with the default observer HUD enabled. Inactive clients select distinct idle sessions, so they do not spectate active ones. This fixes the event fan-out topology. Each active client completes one unscored warmup, then keeps one request outstanding until the common monotonic deadline. Submitted requests drain within their timeout. This measures the wire protocol, excluding browser rendering and external networks.

ACK means admission. Queued requests may emit an empty custody final and a visible follow-up reply under different run IDs, in either order. Unique echoed mock tokens and session identity correlate replies. Empty/status events are not first deltas; duplicate finals cannot advance the loop. Journals retain partial observations and unmatched first deltas by stream identity without inventing a request association.

Results

requests contains send, ACK, first-delta, and visible-final times relative to start. Latencies subtract send time; p50/p95 use nearest rank and include later drained requests. Missing observations remain null. streamObservations retains unbound first-delta identities. Failures, missing replies/deltas, or lost clients invalidate a trial; keep failed trials instead of silently replacing them.

summary.replies and throughput count only fixed-window completions; drainedReplies counts later completions. windowCpu divides fixed-window cgroup CPU by in-window replies, retaining actual boundary times and skew. cpu instead brackets admission through drain and divides by all completions. Both include exited Gateway descendants and exclude drivers/provider; keep denominators separate.

The 50 ms sampler distinguishes lifetime/load-window RSS and thread peaks, recording thread affinity and PID/TID birth ticks. Summed RSS can double-count shared pages and miss brief peaks; it is not cgroup memory.peak. Observations bound thread lifetimes. Verify Gateway JavaScript-worker event coverage separately for each runtime. Workers without individual exit notifications have only the recorded process-exit upper bound; report them separately instead of inventing complete lifetimes. Worker atNs has a runtime-local origin. Use epochMs with the driver's startEpochMs to classify load-window events; raw Node/Bun hrtime origins differ.

Cache policy and proof

Keep user state fresh. Use a separate --transpiler-cache directory per immutable Bun build, prime each shape with one full unscored workload, retain it across scored trials, and record inventories. Empty-cache trials form a separate campaign. Record OS pages, transpiler cache, product-owned Node compile-cache policy, and JIT warmup independently. Compare per-trial summaries from two Node/Bun/Bun/Node blocks per release/shape, including dispersion. Profile separately with exact build IDs and lost-sample counts; individual requests are not independent runtime trials.

Run node --import ./scripts/tsx.mjs scripts/bench-gateway-control-ui-proof.ts /path/to/runtime /delegated/cgroup for opt-in Linux process/protocol proof of exited-child CPU, workers, forced cleanup, reply ordering, disconnects, timeouts, late errors, and cancellation. This stays outside per-PR unit CI. Separately prove a small real Gateway workload's signed admission and subscriptions before scoring. Drivers remain monitored through the coordinated finish/close handshake, with available failure evidence retained. Gateway teardown uses the existing acknowledged IPC stop and shared service stop budget before forced cleanup, preserving admitted descendant work during drain.