Performance

WireGuard performance vs alternatives:

Why WireGuard fast:

• Kernel module: no syscall overhead for packet I/O.
• Single AEAD cipher: no negotiation, no fallback paths.
• Cryptographic routing: no separate ACL lookup.
• Minimal protocol: 4 packet types, no extensions.

Bottlenecks:

• AEAD encryption: ChaCha20 ~3-5 GB/s/core.
• TUN: kernel must copy packets in/out of userspace (for userspace VPN).
• AES-NI hardware: AES-GCM faster on CPUs with AES-NI; ChaCha20 wins on devices without.

Optimizations:

• Multi-queue NIC + RPS: distribute packet processing across cores.
• GSO/GRO: aggregate small packets at NIC level. Saves CPU.
• DPDK / XDP: kernel-bypass packet I/O for max throughput.

UDP fragmentation:

• TUN's MTU should be set ~1420 (1500 - 80 overhead) to avoid fragmentation.
• Don't fragment in inner: kernel drops oversized; ICMP back to source.
• Path MTU discovery via inner packets.

Multi-core scaling:

• Per-peer queues.
• Hash inner packet → CPU.
• Avoid lock contention.

Mobile / low-power:

• WireGuard's small packet count + ChaCha20 win on battery.
• Compared to OpenVPN, much less wake-up time on mobile.