After the NAT two IP (IP1 and IP2) will be IPnat, but the NAT maps the IP1 and IP2 to the port range. Since, there is no port in ICMP, both IP1 and IP2 will be go to uplink as IPg and but on the return there must be problem for NAT machine to traverse the two different paths from IPnat to IP1 and IPnat to IP2. I looked into the ICMP header and observed the packets have different identifiers. So, NAT machine must be using the identifies to reverse the packets.
That's indeed more like a generic ICMP NAT question than a GTP related one. I guess the NAT is going to be using the identifier and changing it to make sure they are unique on the outside world so it can match them when they come back. You can have a look at netfilter stack in the kernel to see what's it doing when NATing ICMP packets.
Anyways, in my case the *IP1=IP2* (In my experimental architecture, the GGSN will not be assigning distinct IP for each host. Instead, GGSN will assign 1 IP address for 32 hosts (seems like NAT). My configuration is probably out of standard architectures, but I need to understand how would gtp handle matching these two pdp contexts. I have tried this configuration, pinging from two different host with same IP and it was successful!
Which GGSN implementation are you using? osmo-ggsn should not allow you to do that afaict. Why are you assigning the same IP to different hosts / pdp ctx ? I'm not sure because I don't know GTP in detail, but from my current knowledge I'd say that this is not a correct setup.
If you plan to have several computers under a given pdp ctx, then use a NAT before sending them to the SGSN, this way the NAT will take care of maintaining correct fields for each ICMP packet coming from different hosts.