Workload-Aware Task Placement in Edge-Assisted Human Re-identification

This work is a cross-domain study by utilizing the most recent cloud and edge computing techniques in the human re-identification, which is a popular computer-vision application motivated by the demand of connecting and monitoring our world in the era of Internet of Things (IoT). We systematically study the real-time re-identification problem within a large-scale video surveillance network. Motivated by the system heterogeneity in terms of real-time workload and hardware configurations, we develop a workload-aware distributed system, which optimally allocates tasks across edge servers and cloud, for pursuing a user-controlled trade-off between system responsiveness & utility. We use an experiment-oriented approach to measure and model the edge heterogeneity. A two-phase task-placement algorithm is proposed which runs with the model built in the off-line phase, and driven by the dynamic real-time workload in runtime. We implement our entire system on a commercial cloud platform and use extensive simulations and experiments to validate its efficacy and responsiveness in practice.