Making Tailscale Faster

Tailscale is delivering a set of data-plane improvements targeting faster, more efficient connectivity on Linux and Android. Work focused first on reducing memory overhead for typical small packets: wireguard-go previously unpacked each small packet into its own 64 KiB buffer, incurring copies and wasted space. New logic leaves many packets inside a single large read, marks their boundaries rather than copying them, and shortens internal packet queues. That change lowers allocations and copy work and yields roughly a 5% speed-up in many configurations. Complementary optimizations include using Linux writev to pass multiple packet pieces to the kernel in one operation, cutting copies and syscalls.

Freed memory enabled architectural changes for high-load nodes: subnet routers, app connectors, and exit nodes move from a single ordered pipeline to a multi-queue model that assigns streams to parallel reader-crypto-writer lanes scaled to CPU cores. That reduces end-to-end delay and raises aggregate throughput, especially for many short-lived connections. Netmap caching lets devices use a cached network map to establish peer connections when the control plane is slow or unreachable (requires a prior successful connect and persistent storage). Memory-buffer changes are planned for v1.104, multi-queue follows, and some throughput gains were partially deployed in spring 2026.