Short version: Adding ITRs, ETRs and a mapping system to
the Internet is adding complexity to the
routing system, but it is not falling into
the trap of the phone system, in which
applications rely on the network being
"smart".
Since it is not practical to solve the
routing scaling problem with host upgrades,
adding a core-edge separation scheme (ITRs,
ETRs and a mapping system) is a way of
providing portability and multihoming to
all end-user networks which want it, while
reducing the burden on routers, and while
maintaining the full Internet service at all
times.
It is adding "stuff" - hardware, software
and a global mapping system - to the Net,
in order to avoid adding something uglier
and more expensive, ever-bigger routers
and a less stable DFZ (while still not
meeting the portability and multihoming
needs of many smaller end-user networks).
In a recent message, Steve Blake mentioned the (a?) guiding principle
of the Internet:
Smart host, dumb network.
This has many benefits over the phone-system, which operates on the
opposite principle. Perhaps OSI involved too much guff in the
network, and during the time the details were being hammered out, the
light and sprightly Internet was born, grew quickly and could be
easily adapted to to Useful things without having to alter the network.
If the Internet had been designed so the hosts were somewhat smarter
- so they could operate perfectly well despite their IP address
changing at any time - then we probably wouldn't have a routing
scaling problem. We would already have something like SHIM6 or ILNP.
If this was true and a 48 bit or 64 bit address space had been
adopted initially, we wouldn't have any shortage of address space either.
As it happens, we have both an IPv4 address shortage and a routing
scaling problem.
I can't see how we could get new host stacks widely adopted or how we
could convince most application writers to do a major rewrite of
their code for the new stack. (Even if we did, users often use old
applications which are no-longer being maintained.)
So I think the only practical approaches to the routing scaling
problem involve adding a new architectural layer to the routing
system: ITRs, ETRs and a mapping system. (AKA a core-edge separation
solution, involving map-encap, forwarding or translation.)
As long as this is done so the new, scalable, form of address space
works well no matter where the hosts are, then this avoids one of the
problems of the "smart network" approach - different physical parts
of it having fundamentally different capabilities and levels of service.
If properly deployed (which depends on business models) Ivip's OITRDs
and I guess LISP's PTRs should enable the new scheme to support all
hosts well enough. However, we need to avoid the situation where a
packet has to travel across the globe to find an ITR which tunnels it
back to near where it came from. Total path lengths through the ITR
and ETR should not generally be much longer than the host-to-host
path. ("Much longer" probably means thousands of kilometres longer.)
With ITRs, ETRs and a mapping system, in principle, as far as host
communications are concerned, the network is still "dumb". Any two
hosts with public addresses can communicate - though distance,
bandwidth restrictions etc. inevitably affect the capacity for
reliable, really fast, communications.
The routers themselves, the Ethernet switches and the fibre links
which make up the Internet may be "dumb" as far as applications are
concerned, but they are certainly not cheap.
Adding ITRs, ETRs and a mapping system to the Net is adding extra
stuff, in order to create a new type of scalable PI space.
I think it is the best option, considering we can't solve the scaling
problem purely with host upgrades. The only other alternative seems
to be not to solve it, and therefore require routers be more and more
expensive, whilst still denying all but the biggest end-user networks
two things which many of them need: multihoming and portable address
space.
The phone system depends almost entirely on the network for the
functionality provided to its users.
Adding ITRs, ETRs and a mapping system to the routing system doesn't
make Internet applications any more dependent on the routing system
than they are today. So it is arguably not making the hosts any
dumber, or the network any smarter.
The network is already exceedingly "smart" inside its components.
Routers (other than those implemented purely in software) - their BGP
implementations and in particular their FIBs - are all intensely
"smart", in terms of the heroics of engineering which are required to
design and manufacture them.
The DNS and the BGP system are big global networks on which
everything depends - but the Internet's routing system is invisible
to applications. I think that adding a core-edge separation scheme
of ITRs, ETRs and a mapping system won't change this.
- Robin
_______________________________________________
rrg mailing list
[email protected]
https://www.irtf.org/mailman/listinfo/rrg
