They do highlight that they use kernel space Wireguard as opposed to Tailscale which uses user space Wireguard.

https://medium.com/netmaker/battle-of-the-vpns-which-one-is-...

https://techoverflow.net/2022/08/19/iperf-benchmark-of-zerot...

Afaik, tailscale recently made changes to their go user space implementation that actually made their version faster than the kernel implementation, at least in some cases.

I remember reading a blog post on tailscacles' website about it and how they are pushing their changes upstream (wg kernel and official wg go user space implementation).

Can't find the post now though.

I think this is the post you're referring to.

https://tailscale.com/blog/more-throughput/

Yes that's the one.

Go is a bit too slow for performant networking - not sure why people are hell-bent on forcing it into in such spaces where it doesn't belong.

This is not true. When it comes to networking, most languages (Go included) are bottlenecked at context switching for syscalls.

Please, elaborate or provide sources.

My hazy understanding from previous conversations was the performance advantages claimed is basically how things were "out of the box" for comparison, for certain situations. However I've seen people claim the differences quickly close when someone that knows what they are doing optimizes a setup. Additionally other situations are much more similar.

The numbers Netmaker posted likely come from: https://medium.com/netmaker/battle-of-the-vpns-which-one-is-...

Note that a spreadsheet with raw data and command used (which is all `iperf3 -c <IP>`) is here: https://docs.google.com/spreadsheets/d/1Qy27zEERSqisdV1u-YUc...

While it's a good idea to be skeptical of a company tooting it's own horn, there does (or has) seemed to be a consistent performance advantage that is beneficial to those who aren't paid for or like being a network admin beyond setting a proper MTU: https://techoverflow.net/2022/08/19/iperf-benchmark-of-zerot...

However in light of https://tailscale.com/blog/more-throughput/ (thanks to FabHK for bringing this to my attention) testing should probably be redone.