End-to-End Performance Analysis of Edge-Based Face Recognition Systems

Abstract

Edge-based video analytics systems are increasingly deployed in public security and access control scenarios. Although accelerator hardware reduces inference latency, end-to-end performance often remains unpredictable. A face recognition service was deployed on a three-node edge cluster equipped with NVIDIA T4 GPUs and 10 GbE networking. The system processed video streams from 96 cameras at 25 fps over a four-week period. Performance metrics were collected at millisecond granularity. Analysis reveals that video decoding and frame transmission accounted for 41% of total latency, while GPU inference contributed only 23%. Storage queue depth exceeded 32 during peak hours, leading to periodic frame drops. After pipeline restructuring, average response time decreased from 184 ms to 127 ms.