Runtime telemetry from evolutionary and static resource policies on a Jetson Nano

Abstract We describe 108 hours of recorded execution on an NVIDIA Jetson Nano (2 GB): eighteen six-hour missions in which four image-processing workloads alternate in twelve 1 800-s phases. Six missions run a persistent surrogate-assisted evolutionary runtime controller with adaptive objective weights and a thermal supervisor, two under each of three application thermal profiles; twelve run fixed CPU-only, GPU-only and balanced policies on the same schedule. Every mission carries a 2-Hz system trace, one record per finalized camera frame (8.08 million in all), an external power-meter stream, a hot-swap log and a sealed collector event log; controller missions add one record per 800-ms control period (161 723) with requested controls, software read-back and the frozen forecast of each decision, retrospectively associated with recorded outcomes along the selected trajectory. The two missions per arm and profile are chronological repetitions, not randomized replicates. The recordings separate requests, read-back and measured outcomes, were validated channel by channel, and are released with verifiers and the scripts that regenerate every derived table.