Design and Implementation of a Traffic-Aware Adaptive Flow Offloading Strategy on OVS-DPDK with Hardware Acceleration

Abstract

As cloud data centers increasingly adopt software-defined networking (SDN), network function virtualization (NFV), and hardware offloading, virtual switch packet processing efficiency has become critical. Although Open vSwitch (OVS) combined with the Data Plane Development Kit (DPDK) improves forwarding performance by bypassing the kernel network stack, the hardware flow table of SmartNICs remains a limited resource that should be used efficiently. Many existing OVS-DPDK hardware offloading strategies rely on fixed offloading limits and cannot dynamically adapt hardware resource usage to real-time traffic changes. This paper proposes a Traffic-Aware Adaptive Flow Offloading (TA-AFO) strategy integrated into OVS-DPDK and deployed on an NVIDIA BlueField-2 DPU. TA-AFO shifts the offloading trigger to fast-path flows whose cumulative hit count exceeds a dynamic threshold, ensuring only long-lived flows occupy hardware flow table resources. The threshold is continuously adjusted based on OVS-DPDK’s real-time processing load, enabling automatic adaptation to changing traffic patterns. Experimental results show that, under simulated data center traffic, TA-AFO keeps hardware flow-table occupancy at approximately 2.1% of the observed datapath flows (97.9% reduction relative to offloading all flows), maintains throughput around 4.4-4.6 Mpps in steady state, and keeps packet drop rates near zero across most of the experiment duration.