Systems Networking Architecture
The latency/throughput ladder (pick the lowest tier that meets needs)
| Tier | Mechanism | Throughput/latency | Cost/complexity |
|---|---|---|---|
| Sockets + epoll | Classic kernel stack | Good; syscalls + copies dominate | Lowest |
| io_uring | Async, batched, shared rings | Fewer syscalls/copies | Moderate; Linux ≥5.x |
| eBPF/XDP | Run code at the driver hook | Drop/redirect at line rate before stack | Verifier limits; per-packet logic |
| Kernel bypass (DPDK) | Poll-mode driver in userspace | 10–100M pps, µs latency | High; burns cores, no kernel stack |
Decision drivers: required pps/latency, CPU budget (DPDK busy-polls cores), need for the kernel stack (TLS, routing), and operational complexity.
Architectural levers
- Zero-copy: avoid user/kernel copies (sendfile, splice, io_uring fixed buffers, DPDK mbufs). Copies are often the real bottleneck.
- Per-core / shared-nothing: thread-per-core with RSS/affinity (Seastar model) to kill cross-core contention and cache bouncing.
- NIC offloads: checksum, TSO/LRO, RSS, and increasingly full TCP/crypto offload; SmartNIC/DPU pushes the data plane off the host CPU.
- Congestion control: choose per goal — CUBIC (throughput), BBR (latency/bufferbloat), DCTCP (datacenter). Matters for tail latency.
- C10M: the bottleneck is per-connection kernel cost and copies → bypass + zero-copy + per-core state.
When to recommend what
- API server, normal scale → sockets/epoll or io_uring; don't over-engineer.
- Software router/firewall/LB at line rate → XDP/eBPF or DPDK.
- µs-tail trading/telco data plane → DPDK + per-core + kernel-bypass, SmartNIC.