15.253 14.9.13 Perfiles y presupuestos de recursos
| ID | Perfil | Prioridad | Concurrencia | Uso |
|---|---|---|---|---|
| RES-0 | AcquisitionSafe | Capture > metadata commit > health > inference | Inference paused or one low-priority worker | Critical storage/thermal incidents |
| RES-1 | OrinNano8GBBalanced | Capture reserved; one GPU-heavy stage at a time | Bounded CPU workers; GPU serialization unless benchmark proves overlap | Recommended initial MVP |
| RES-2 | OrinNano8GBBatchQuality | Capture plus deferred high-quality offline analysis | Batch after attempt/session; aggressive cache eviction | Maximum local quality without real-time claim |
| RES-3 | WorkstationDevelopment | Debug fidelity and parallel experimentation | Hardware-dependent | Development only; not field SLO |
| RES-4 | CPUFallback | Capture and minimal analysis | Serial/small models | GPU unavailable or compatibility fallback |
- BUD-01: Reserve memory and IO headroom for capture before allocating inference workers.
- BUD-02: Set explicit container/service memory limits only after accounting for shared GPU memory behavior on Jetson.
- BUD-03: A process killed by OOM is a failed job, not a transparent retry without profile change.
- BUD-04: GPU engines are cached by exact model, TensorRT/runtime, precision, shapes and hardware profile.
- BUD-05: Use bounded frame queues; choose DropOldest, DropNewest or BlockCapture only through an explicit stream policy.
- BUD-06: Required evidence streams may not use silent drop policies.
- BUD-07: Batch size and resolution are configuration, telemetry dimensions and benchmark variables.
- BUD-08: Thermal and power mode are part of the deployment profile and benchmark evidence.
- BUD-09: Use tegrastats/OS metrics to measure memory, processor, temperature and power-related state on Jetson.
- BUD-10: Schedule heavy reprocessing outside capture windows when concurrent execution violates the quality floor.
- BUD-11: Measure encode/decode and storage IO separately from neural inference.
- BUD-12: Use a latency budget tree; each stage owns a budget and queue delay is reported separately.
- BUD-13: Do not compare workstation latency with Jetson latency as if they were the same deployment track.
- BUD-14: Any dynamic degradation emits a profile-change event and remains visible in the result provenance.
tegrastats reporta uso de memoria y procesadores y puede
integrarse como fuente local. Las métricas derivadas deben conservar
versión de parser porque el formato puede variar entre releases
[P49-R10]. Los modos de potencia configurados mediante
nvpmodel forman parte del perfil, no de una optimización
invisible [P49-R11].
15.253.1 Árbol de latencia
time_to_result =
capture_finalize
+ queue_wait
+ decode/preprocess
+ pose
+ scene
+ contact/phase/features
+ classification
+ quality/assembly
+ publishCada sumando se mide por separado. Un throughput alto con queue wait creciente no se reporta como sistema saludable.