An implementation of this paper: A New Method for Symmetric NAT Traversal in UDP and TCP

According to RFC 3489, a symmetric NAT is one where all requests from the same internal IP address and port, to a specific destination IP address and port, are mapped to the same external IP address and port. If the same host sends a packet with the same source address and port, but to a different destination, a different mapping is used. Furthermore, only the external host that receives a packet can send a UDP packet back to the internal host.

So, if we know the port allocation rule of the Symmetric NAT, we can traverse Symmetric NAT. This paper proposes a new method for traversing Symmetric NAT which is based on port prediction and limited TTL values.

This method is based on limited TTL values, port prediction(if it's not predictable, use large number of holes, namely a 1000 connections at once, which will be punched and increase the success rate).

It's just an experimental project, I just wanna test whether UDP punching is possible if both nodes are behind symmetric NAT. It works this way:

A peer got its own NAT info(external ip/port, NAT type ), then registers to server, the server replies to the peer with a unique peer ID. if you specify -d peer ID option, the node tries to get the info of that specified peer from the server, then connects to it directly. So when you specify '-d' option, make sure that peer is connected with server.

To run it, you should run punch_server.go in a machine with public IP, so that both nodes can connect to it to exchange info(just node info, not to relay payload like TURN), then run nat_traversal [-s punch server] [-d if you wanna connect to peer] on clients, you can also specify other arguments, such as, STUN server by -H option, source IP by -i, source port by -p.
This program is not 100% guaranteed to make 2 peers behind symmetric NATs connect to each other, it's possible in theory. Actually, I just happen to succeed once. Hope more people can test it.