Re:A crisis in the making: Only 4% of the Internet supports ...
by tptacek on Friday March 06, 2009 @03:05PM
(#505883)
I think that one of the major architectural failings of the IETF over the past N years is the belief that we should be building on and extending the core Internet protocols, when what we should be doing is moving more and more logic into the application layer.
So for instance, if you want to be able to address billions of people, the scheme that you use to do that should look more like email addresses (which are effectively unbounded) and less like a globally synchronized database of integers.
A lot of problems that are basically intractable at the router/switch layer are much easier to solve with relay servers. Multicast is another great example; IP multicast was a failure, but multicast delivery semantics with (say) 80% of the performance improvements of IP multicast are achievable with relays.
This isn't a crackpot idea; Frans Kaashoek and rtm at MIT had a project investigating it, called Reliable Overlay Networks. I did a startup in the late '90s based on overlays. The technology worked great.
The pain involved in NAT and 1918 addresses would be mitigated if our applications were less tightly coupled to IP (gethostbyname -> sockaddr_in -> connect), and if we could just pick a NAT traversal / NAT rendesvouz technology and run with it.
So my answer to your question is, you shouldn't worry about the alternative. Just keep building normal socket applications.
So for instance, if you want to be able to address billions of people, the scheme that you use to do that should look more like email addresses (which are effectively unbounded) and less like a globally synchronized database of integers.
A lot of problems that are basically intractable at the router/switch layer are much easier to solve with relay servers. Multicast is another great example; IP multicast was a failure, but multicast delivery semantics with (say) 80% of the performance improvements of IP multicast are achievable with relays.
This isn't a crackpot idea; Frans Kaashoek and rtm at MIT had a project investigating it, called Reliable Overlay Networks. I did a startup in the late '90s based on overlays. The technology worked great.
The pain involved in NAT and 1918 addresses would be mitigated if our applications were less tightly coupled to IP (gethostbyname -> sockaddr_in -> connect), and if we could just pick a NAT traversal / NAT rendesvouz technology and run with it.
So my answer to your question is, you shouldn't worry about the alternative. Just keep building normal socket applications.