Honeybee-DR: dynamic dependency management and lightweight reliability for mobile crowd computing

Opportunistic mobile clusters formed by smartphones and tablets in disaster-response, field-science, classroom, and ad-hoc multi-camera settings can run demanding workloads, but only if two problems are addressed: the workflow must adapt to runtime conditions rather than being fixed in advance, and worker results must be reliable without paying the cost of full replication or cryptographic verification. This paper presents Honeybee-DR, a delegator-centric framework that addresses both problems. The dependency-management mechanism centralizes orchestration logic on the delegator and supports three patterns of runtime DAG growth (data-dependent branching, iterative refinement, conditional cascade) while keeping workers stateless. The lightweight reliability mechanism injects adaptive trap jobs through the same JobPool as ordinary work, updating each worker's score with an asymmetric, forgiving rule that tolerates transient faults while concentrating validation on persistently unreliable workers. On a heterogeneous 5-device Android testbed running a multi-camera video-processing workload, Honeybee-DR (Dynamic) reduced processing time by an average of about 17%, delegator energy by 21%, and peak delegator memory by 30.5% versus its static counterpart, with a discovery overhead of 96.5 ms per video segment. Under a face-detection workload with two rogue workers, the adaptive trap-job mechanism reached 99.5% accuracy with up to 80.5% lower validation overhead than majority-vote replication.