Magic sockets via LD_PRELOAD?

Question

Computer networks

I read this article on Habré. There is a comment

Melanox and Ethernet can speed up: they showed me tests with “pings”, 9.5 µs on 40G cards, it seems without a switch. After replacing standard sockets with "magic" sockets through LD_PRELOAD, 1.5 μs is obtained, while the entire tcp / ip stack is taken over by the network. That is, on each machine, 4 μs of overhead was removed.

What are these "Magic" sockets? How is this done (software or hardware) and is it possible to repeat this without the appropriate technologies from melanox, with conventional network cards? In which direction to dig?

Answer 1

I mean this https://github.com/Mellanox/libvma/wiki .
The googling word is RDMA (well, and around it).
The bottom line is that this requires special hardware, so as far as I know, nothing will come of it on a "regular network card".