approaches

DPDK, XDP (using eBPF in kernel)

Solarflare TCPDirect (L3/4 network stack)

OpenOnload (no change required, just need to attach a shared libraries → will detect UNIX network syscall and override…)

Xillinx

https://blog.cloudflare.com/kernel-bypass/

https://lambdafunc.medium.com/kernel-bypass-techniques-in-linux-for-high-frequency-trading-a-deep-dive-de347ccd5407

Solarflare NIC can connect with any server motherboard with standard PCIe slot

vs io_uring

kernel bypass is faster:

• zero overhead:
  • if use io_uring, packets still goes into the kernel through routing tables, firewall rules (iptables/nftables), socket buffer management (sk_buff), and security checks
  • io_uring still requires kernel to manage page tables to store data and map between kernel space and user space…!
  • Kernel bypass uses DMA (iommu, vfio) to write straight to application memory
• kernel scheduling:
  • io_uring depends on kernel to schedule background helper threads (SQPOLL, io_wq) → if these threads are migrated to another CPU core due to heavy load → cache locality lost → spike!
  • kernel bypass tends to isolate CPU (only DPDK or????) → no deschdule, etc. → better perf
• cache locality:
  • io_uring jumps between application threads and the kernel's asynchronous workers → invalidate caches…
  • kernel bypass executes the hardware driver logic inside the exact same CPU instruction loop as the application logic, the data remains entirely within the CPU's local L1/L2 caches.