← Back to the chapter · Next detail →

Technical detail · Fog Energy-Latency Tradeoffs

What MAUI and LEO add to the ledger

MAUI profiles each mobile code component’s energy cost and decides at runtime whether to execute locally or offload. It includes the energy and latency cost of moving code and data, not just running them. The decision is per component and network condition rather than one always-local or always-cloud setting for the application.

LEO is a low-power scheduler for continuous and triggered mobile sensing, including accelerometer-based activity recognition and audio tasks such as keyword spotting, emotion recognition, and speaker identification. Low-overhead heuristics run on a dedicated low-power unit and assign sensor-processing tasks to the CPU, DSP, or cloud according to device energy and latency constraints; the paper also evaluates a GPU. Scheduling five apps once per second took 107 ms per schedule. The paper estimates that on-demand scheduling at that rate consumes less than 0.5% of a 2300 mAh battery per day (Georgiev et al., LEO).

The study’s optimized on-device GPU keyword-spotting pipeline returned results about 6x faster than a 5 Mbps cloud path, while batched GPU scheduling could use less energy than 1 Mbps cloud offloading (Georgiev et al., LEO). These results describe the tested mobile GPU, workload, and network baselines. Re-measure on the actual hardware and radio before repeating them.

Both systems retain the local-compute-versus-transfer comparison. They add a runtime that keeps measuring and deciding rather than fixing placement once. The scheduler itself must pay for its energy and time overhead.

← Back to the chapter · Next detail →