I developed a C++ library which implements AF\XDP and DPDK as backends for latency critical network processing tasks. I notice there are only few benchmarks comparing AF_XDP against DPDK on the same NICs therefore I aim to fill this gap. I used C++20 since that is my strongest language and it interfaces nicely with the low-level driver or linux kernel code.
The AF_XDP implementation is here: https://github.com/ASherjil/ABTRDA3/blob/master/src/backends/AF\_XDP/AFXDP.hpp
The comprehensive benchmark 24h results are here:
https://github.com/ASherjil/ABTRDA3/blob/master/docs/Benchmarks.md
One point that I think is underappreciated: a commonly marketed advantage of AF_XDP is that it doesn't unbind the NIC driver, so the interface stays visible to
ip link and ethtool. That's true for Intel NICs, where DPDK requires binding to vfio-pci. But it's not true for mlx5 (ConnectX-4/5/6/7), which is a bifurcated driver — the NIC stays fully visible to the kernel while DPDK runs. On Mellanox hardware, the main practical argument for AF_XDP largely disappears.On implementation effort: even with
libbpf and libxdp, custom AF_XDP has high code complexity and plenty of driver-specific quirks. Its four lock-free SPSC rings (fill/RX/TX/completion) are counterintuitive and fairly difficult to comprehend. Without the helper libraries I'd call it a serious long-term project. For most use cases I'd recommend DPDK's AF_XDP PMD instead of a from-scratch implementation — in my results, my custom implementation performed only marginally better than DPDK’s AF_XDP PMD. I've documented the driver specific issues in the repository. Let me know if you guys have any feedback about the code, bench-marking methodology or anything else. Or any questions I'm happy to answer I don't really mind.
https://redd.it/1uub2kx
@r_cpp
