Impact Analysis of NUMA-Aware Flow Placement Based on Network Flow Size in Multicore Systems

This paper analyzes the impact of Direct Cache Access (DCA) within a multicore Non-Uniform Memory Access (NUMA) framework on network packet processing efficiency in Linux systems. By strategically placing network flows based on their characteristics and NUMA locality, we examine the performance dynamics of integrating and distributing long and short flows across different NUMA nodes. Utilizing Advanced Receive Flow Steering (aRFS) to differentiate and steer flows, we provide insights into the benefits of leveraging DCA for long flows and using DMA for short flows. Our study reveals significant improvements in network throughput by reducing data copying overheads, highlighting optimized traffic management strategies for high-demand data centers.