Evaluation of Scaling Method for Frame Reception Processing of Media Stream Using XDP/AF_XDP on Edge Device

In recent years, edge AI, which performs low-latency inference processing near IoT devices such as network cameras, has become increasingly popular among consumers. In the future, with improvements in network speed and inference processing throughput on edge devices, there will be a need for scalable processing capabilities throughout the entire frame processing pipeline, from media stream packet reception to inference processing. Receive-Side Scaling (RSS) is commonly used as a method to scale the processing load of packet reception across multiple cores. However, since media such as video is often delivered as a single stream, packets cannot be distributed based on sessions like IP addresses or ports. Therefore, a mechanism is needed to distribute packets to each CPU core using other elements, such as frame numbers. Recently, SmartNICs, which allow the implementation of custom logic to analyze packet content and distribute it to CPU cores, have become widely used. However, for edge devices processing specific streams, SmartNICs are often too costly and over-specification. This paper presents a low-cost and low-latency method for scaling frame reception processing for single-stream packet reception using Linux eXpress Data Path (XDP) and the Address Family (AF_XDP). Our results indicate that while keeping the load of sequential processing constant, the overall load is reduced by 57% with 10 cores.