In this case I cannot map the IP address of the TCP session to one specific
mobile subscriber - and the only way I can
identify the subscriber is by "looking" on the SMPP layer (above the TCP)
and extract the subscriber mobile number.
It's tempting to suggest using IPv6 tunelling, and sorting out whos packets
are whose by inserting the subscriber info in the low-order 64 bits.
I would love to see the neighbor discovery (~= ARP for the v6-challenged) for such a solution. Not to mention that you can't run TCP that way.
But why use TCP in the first place? If this stuff is small enough to fit in a single IP packet, it makes much more sense to use UDP: no multiplexing problems. If reliability is needed it gets more complex, but some kind of RTP-like protocol makes more sense here than TCP.