Benchmarks & Profiling
Lifecycle transition latency, adapter load time, and control-plane overhead
Measurements in Preparation
Formal real-device benchmark measurements for AMEVA-Orchestrator are being prepared across the 5-device fleet. The orchestrator's own overhead is typically sub-millisecond; meaningful measurements require full adapter lifecycle sequences (COLD→HOT→ACTIVE→COLD) on real modality engines.
Planned measurement axes: adapter lifecycle transition latency (ms, all 5 tiers), UDP PIN pairing handshake time (ms), SHA-256 audit-chain append latency (µs), secret redaction throughput (MB/s), and total RSS overhead of the control-plane process (MB).
Planned Measurement Conditions
| Metric | Planned Device | Conditions | Status |
|---|---|---|---|
| COLD → HOT lifecycle transition | Galaxy S25 / A35 (Termux) | LLM adapter (Qwen2.5-0.5B Q4_K_M), cold start | Pending |
| UDP PIN pairing handshake | S25 (master) + A35 (worker) | Tailscale mesh, 6-digit one-time PIN, HMAC beacon | Pending |
| SHA-256 audit-chain append | Galaxy S25 | 10K consecutive command entries, single-threaded | Pending |
| Secret redaction throughput | Galaxy S25 | 1 MB log stream, Bearer+API-key+URL-credential patterns | Pending |
| Control-plane process RSS | Galaxy S25 / A35 | 9 adapters registered, 3 active, steady-state | Pending |