Poster Abstract: Characterizing Computational Workloads in UAV Applications

Abstract

Unmanned aerial vehicles (UAV) enable novel but demanding computational workloads that exceed processing capacity of their onboard resources. Mobile and edge devices can support demanding workloads, but they increase network communication and power usage. These resource constraints block potentially transformative UAV applications that execute too slowly or use too much power. This poster presents early efforts to characterize computational demands of emerging UAV applications. We are building a selfie-drone benchmark. Our benchmark will capture processing metrics, e.g., CPU usage, cache misses, power usage, etc. It will also enable characterization across a wide range of local and edge setups. Our benchmark uses a micro services design, making it easy to move workload execution across multiple contexts. Early results show that our benchmark functions well, i.e., accurately detects faces and safely uses flight controls. Further, edge devices matter. A smart phone uses 4X less power than a laptop when executing our benchmark.