Huh, would be interesting to see some non-tailscale benchmarks of this. Assuming the kernel impl is actually optimized it should be theoretically impossible to exceed the performance with userland wg?