The Blind and the Elephant: A Preference-aware Edge VideoAnalytics Scheduler for Maximizing System Benefit

Abstract

Video analytics is the killer workload in edge computing, which in-volves the scheduler’s complex decisions to balance analysis perfor-mance (latency and accuracy) and resource consumption (network,computation, and energy). Traditional schedulers address this as asingle-objective optimization problem with fixed weights, unableto precisely capture unknown system preferences due to intricatepricing rules across various service levels and resource costs, con-sequently leading to suboptimal system benefit like monetary gain. In this paper, we propose a Bayesian optimization-driven multi-objective scheduler, PaMO, that can proactively explore the systempricing preference by pairwise comparing outcome vectors of all ob-jectives. Moreover, PaMO designs a heuristic scheduling algorithmwith a zero-delay jitter guarantee to avoid performance degrada-tion caused by resource contention and uses a revised Bayesianoptimization algorithm to make video configuration and scheduling decisions. Experiments on real video analytics workloadsshow that PaMO can achieve up to 53.9% benefit gain compared tostate-of-the-art scheduling methods.