Open vSwitch with AF_XDP as a userspace datapath
summary
Open vSwitch implements an AF_XDP userspace datapath and measures it with OVS perf pvp tests and physical loopback. Results: on physical loopback AF_XDP hits 8.8M pps vs 5.6M pps for the kernel datapath, but through a VM AF_XDP is about 1.6M vs 2.75M through VM for DPDK. Kernel datapath can use ~8–9 cores; AF_XDP shows higher total CPU due to extra pollers. Implementation works but lacks features (QoS/egress policer), has buffer-copy overhead, and zero-copy work is ongoing.
AI-written summary · about
chapters (15)
- 0:00Intro
- 0:13AFXDP
- 3:20Comparison
- 4:11Results
- 6:17FXVP test
- 7:20PVP test
- 7:58Frame size comparison
- 8:37Performance comparison
- 10:00Conclusion
- 13:22AFXDP vs DPDK
- 13:57What is the main advantage
- 15:25Reused buffers
- 15:47Future work
- 16:28CI testing
- 16:41Conclusions
related talks
20mCan OVS-DPDK be further optimised?
14mPerf Characteristics of Userspace & Kernel-Space Datapaths in Container Networking
23mMigrating from DPDK to AF_XDP for High-Performance Networking in...
25mDPDK performance optimization for edge computing scenario
37mEfficient and Flexible Virtual Machine Networking Through eBPF
28mLow-Latency, Deterministic Networking with Standard Linux using XDP Sockets
28mXDP for TSN
28mDPDK vs ODP: A comparison