XDP for TSN
summary
Measured XDP/AF_XDP bypass reduces packet path latency from prior ~60–70 µs to under 10 µs end-to-user‑space; a 10,000‑packet test at 1,000 PPS shows ~3 µs average and 16 µs peak, and disabling MMIO moves many samples below 2 µs. Tests run on Cortex A53 and an AM572 Cortex A15 with a 4.19 RT kernel, comparing AF_XDP vs AF_PACKET and poll vs system wake; AF_XDP zero-copy unsupported on tested NICs, polling (interrupts off) drops latency but pins CPU.
AI-written summary · about
chapters (12)
- 0:00Intro
- 1:49XDP Overview
- 3:44FXDP Overview
- 5:42Why XDP for TSN
- 8:26BPF Program
- 11:30Initial tests
- 13:16Latency graph
- 14:43Test model
- 15:57Latency comparison
- 18:27Software latency
- 20:54Jitter
- 25:14Latency
related talks
28mLow-Latency, Deterministic Networking with Standard Linux using XDP Sockets
22mExtending AF_XDP with hardware metadata
33mXDP gaining access to NIC hardware hints via BTF
18mOpen vSwitch with AF_XDP as a userspace datapath
21mHow to Use XDP and eBPF to Accelerate IPSec Throughput by 400%
33mBringing packet queueing to XDP
28mKNOD: In-Kernel Network Offload Device for GPU-Accelerated Packet Processing
19mGet start with XDP: write a network filter in 10m