[vxfemboy]

Resolves #173

Summary

This PR addresses Issue #173 by:

1. Making MAP_POPULATE configurable in UmemPool:

   • Added UmemPoolConfig struct controlling chunk count, chunk size, page pre-faulting (pre_fault: bool), and physical page locking (lock_memory: bool via libc::mlock).
   • Added UmemPool::new_lazy and UmemPool::with_config allowing on-demand page faulting, dropping initialization latency from 1,590.98 µs to 7.25 µs (219.5× faster startup) while reducing resident set size (RSS) on memory-constrained systems.
   • Added UmemPool::new_slabs_with_config for partitioned slabs.
   • Retained full backward compatibility with default UmemPool::new pre-faulting for zero-jitter bare-metal deployments.
   • Added runtime environment configuration in bin/yipd/src/sharding.rs via YIP_UMEM_LAZY and YIP_UMEM_MLOCK.

2. Expanding BatchUdpSocket Syscall Burst Sizing:

   • Generalized recvmmsg_batch<const N: usize> and sendmmsg_burst<const N: usize> using const-generic stack allocations with 0 heap churn.
   • Added BATCH_SIZE_64 = 64 and BATCH_SIZE_128 = 128 constants alongside default BATCH_SIZE = 32.
   • Preserved sendmmsg_batch delegating to 32-packet bursts for zero breaking changes.

3. Multi-Region AWS WAN Benchmarking & Graph Updates:

   • Executed live best-of-3 WAN benchmark runs between AWS us-east-1 and us-west-2 across baseline, Yip, WireGuard, and OpenVPN.
   • Reduced into crates/yip-bench/wan/results/best_of_3/wan.csv and updated crates/yip-bench/RESULTS.md and README.md.
   • Regenerated benchmark SVG visualizations (assets/bench-throughput-loss.svg and assets/bench-fec-recovery.svg).

Microbenchmark Results (af_xdp_scale)

Metric / Component	Baseline (MAP_POPULATE)	Lazy On-Demand (new_lazy)	Improvement
UMEM Init Latency	1,590.98 µs/pool	7.25 µs/pool	219.5× faster startup
Syscall Burst 32	0.210 Mpps (4,766 ns/pkt)	0.210 Mpps	Baseline
Syscall Burst 64	—	0.208 Mpps (4,805 ns/pkt)	2× datagram payload/call
Syscall Burst 128	—	0.217 Mpps (4,616 ns/pkt)	4× datagram payload/call
Packet Drops / Reorder	0 drops / 0 OOO	0 drops / 0 OOO	Strict monotonic FIFO

Real-World AWS WAN Benchmarks (us-east-1 ↔ us-west-2, Best of 3)

Two 1-vCPU Debian 13 VPSes, ~54.365 ms RTT apart, 2026-10-27.

Clean-Path Throughput & Latency

Tunnel	TCP 1-stream	TCP 4-stream	UDP Goodput	UDP Loss	Idle RTT	Under Load	CPU (Host A)
Baseline (no tunnel)	390.7 Mbit/s	906.1 Mbit/s	200.0 Mbit/s	0.00%	54.365 ms	54.360 ms	1.7%
yip	253.7 Mbit/s	350.7 Mbit/s	200.0 Mbit/s	0.00%	55.288 ms	56.051 ms	12.5%
WireGuard	453.1 Mbit/s	1077.7 Mbit/s	200.0 Mbit/s	0.00%	54.981 ms	55.067 ms	11.2%
OpenVPN	32.6 Mbit/s	35.6 Mbit/s	200.0 Mbit/s	0.04%	54.745 ms	54.661 ms	3.7%

Loss Resilience (Injected Outer Egress Loss)

Injected Loss	Baseline Goodput	WireGuard Goodput	OpenVPN Goodput	yip Goodput	yip Effective Loss
0%	497.9 Mbit/s	455.6 Mbit/s	27.7 Mbit/s	307.2 Mbit/s	0%
1%	32.8 Mbit/s	2.1 Mbit/s	3.7 Mbit/s	310.5 Mbit/s	0%
3%	2.2 Mbit/s	1.3 Mbit/s	1.2 Mbit/s	179.3 Mbit/s	0%
5%	1.6 Mbit/s	1.2 Mbit/s	1.0 Mbit/s	232.6 Mbit/s	0%
10%	0.7 Mbit/s	0.6 Mbit/s	0.6 Mbit/s	44.2 Mbit/s	0%

Verification

• cargo test --workspace: All 391 unit + 98 integration tests pass cleanly.
• cargo clippy --workspace --all-targets -- -D warnings: 0 warnings, 0 errors.
• cargo fmt --check: 0 diffs.
• cargo shear: 0 unused dependencies.
• python3 crates/yip-bench/generate-graphs.py --check: OK.
• python3 crates/yip-bench/generate-wire-format.py --check: OK.
• pytest crates/yip-bench/: 16 passed in 0.17s.