Re:Exploring sub-50ns NIC receive paths via userspace MMIO
by krish678 on Tuesday December 16, 2025 @03:30AM
(#46286137)
Hi HN,
I’ve been experimenting with how far NIC receive latency can be pushed on
commodity hardware when you bypass the kernel and treat the NIC as
memory-mapped hardware instead of a network device.
This is a research / measurement project, not a complete networking stack.
What I’m measuring:
• ~20–50ns NIC RX → userspace visibility (device-dependent)
• Busy-polling descriptor rings via MMIO
• Zero-copy DMA buffers
• No syscalls, interrupts, or kernel networking
Important clarification:
The public code focuses on instrumentation, register access, and timing
methodology. Some experimental logic and tuning paths are intentionally
not published yet.
This is mainly about how to measure and reason about latency accurately
at this scale, not claiming a drop-in replacement for DPDK or kernel stacks.
Questions I’d genuinely like feedback on:
1. Where do you usually see the first sources of jitter below ~100ns?
2. On ARM64, how much penalty do you typically expect from memory ordering?
3. Any recommended techniques for validating RX timestamps externally?
I’ve been experimenting with how far NIC receive latency can be pushed on commodity hardware when you bypass the kernel and treat the NIC as memory-mapped hardware instead of a network device.
This is a research / measurement project, not a complete networking stack.
What I’m measuring: • ~20–50ns NIC RX → userspace visibility (device-dependent) • Busy-polling descriptor rings via MMIO • Zero-copy DMA buffers • No syscalls, interrupts, or kernel networking
Public repo: https://github.com/krish567366/BareMetalNIC
Docs / notes: https://baremetalnic.krishnabajpai.me
Important clarification: The public code focuses on instrumentation, register access, and timing methodology. Some experimental logic and tuning paths are intentionally not published yet.
This is mainly about how to measure and reason about latency accurately at this scale, not claiming a drop-in replacement for DPDK or kernel stacks.
Questions I’d genuinely like feedback on: 1. Where do you usually see the first sources of jitter below ~100ns? 2. On ARM64, how much penalty do you typically expect from memory ordering? 3. Any recommended techniques for validating RX timestamps externally?
Happy to clarify methodology or measurements.