Skip to content

performance.md

Latest commit

 

History

History
76 lines (57 loc) · 3.61 KB

performance.md

File metadata and controls

76 lines (57 loc) · 3.61 KB
Raw

Runtime performance and Web UI bundle tracking

Run commands from the Secutils root. Retrack owns its own harness and guide. Read source scenario definitions before interpreting results because runtime architecture changes can alter their meaning.

JS runtime harness

The harness in benches/js-runtime-perf/ links the real secutils::js_runtime::JsRuntime. Read src/js_runtime.rs, src/js_runtime/worker_pool.rs, and src/js_runtime/op_proxy_request.rs for the current execution architecture. Workers retain Tokio runtimes and proxy clients. Script executions create fresh V8 isolates. The cold-start scenario name does not establish that every execution creates a thread or HTTP client.

Use make help and the harness CLI for flags/defaults. Read benches/js-runtime-perf/src/scenarios.rs for default and on-demand scenarios.

make perf
make perf ANALYZE=1 PERF_ITERATIONS=20 PERF_WARMUP=5
make perf PERF_OUTPUT=/tmp/perf-baseline.json
make perf-analyze
make perf-report

make perf produces the report without history changes. ANALYZE=1 compares against .perf/history.jsonl and can append a history entry. When comparison/history recording is intended, use ANALYZE=1. Benchmark regressions produce advisory warnings. Execution or analysis failures still require investigation. Compare platform-matched results through env.os, env.arch, and env.cpuModel.

Resident isolates

resident_isolates is a manual capacity probe outside the default ALL scenarios and tracked history. Name it explicitly and divide peak_rss_delta_kb by concurrency for marginal resident cost. Use RESIDENT_PARK_MS to extend the park window for large concurrency. Read the scenario for its default and clamp.

cargo run -p js-runtime-perf -- --scenarios resident_isolates --concurrency 1000
SECUTILS_JS_WORKERS=1000 SECUTILS_JS_MAX_WORKERS=1000 RESIDENT_PARK_MS=12000 \
  cargo run --release -p js-runtime-perf -- --scenarios resident_isolates --concurrency 1000

The second command prewarms workers to distinguish isolate cost from thread startup.

Scenarios and history

For a default scenario, register its module, run dispatcher, ALL entry, and .perf/config.json entry. For an on-demand probe, use ON_DEMAND and the dispatcher, outside ALL and tracked configuration. Completion requires the scenario to run and produce the expected report fields.

Read scripts/analyze-perf.ts for tracked metrics and HISTORY_APPEND_THRESHOLD_PCT. History appends only on material movement, scenario changes, or structural zero/nonzero transitions. Read .perf/config.json for advisory thresholds. Positive latency/RSS deltas and negative throughput deltas indicate regressions.

CI records reports and can commit material history changes with [skip ci]. The root perf and bundle jobs share dev/scripts/commit-ci-history.sh and can race. Keep each job's history file distinct and append-only so fetch/rebase/retry remains valid.

Web UI bundles

Before bundle analysis, build Web UI output:

npm --prefix components/secutils-webui run build
npm --prefix components/secutils-webui run analyze

build:analyze also cleans before build and analysis. The analyzer reads dist/, compares tracked bundles, and appends material history changes. Regressions are advisory. Missing output is an analysis failure.

For a new lazy-loaded workspace route, add its pattern to components/secutils-webui/.bundlesize/config.json. Otherwise, its output falls into untracked other chunks. Completion requires the new bundle to appear in the tracked analysis output. Read the configuration and analyzer for budgets, patterns, and append thresholds.