Re: geo short vs long term? [Re: Geo pros and cons]

[Iljitsch van Beijnum <iljitsch@muada.com>]

On donderdag, apr 3, 2003, at 22:09 Europe/Amsterdam, Tony Li wrote:

> - The Internet is continually growing at an exponential rate.
>   Most people seem to peg the growth rate at 100% per year
>   currently.  The exact number is not an issue.
Agree.

> - In the past, we've estimated that 10% of all sites would
>   multi-home.  Let's assume a constant rate of 10% of the
>   world is an exception to the default aggregation rules
>   that we pick.
> - From the above two, we can reason that our exception rate
>   is going to continue to grow exponentially.  Note that
>   the rate of absolute growth is more of an issue than the
>   exception rate.
> - Moore's law for memory suggests that memory sizes will
>   double about every two years.  However, memory speeds will
>   not keep up.
So far so good.

> - Packet lookups are a function of memory bandwidth, so to
>   sustain Internet bandwidth growth of 100% per year, we need
>   to also increase memory bandwidth by about 100% per year.
>   Using bigger, slower memories is not a realistic option.
Am I missing something here, because I always thought looking up routes 
scales O(log(n)). If memory size isn't a problem, you could even use a 
2^45 element array and make route lookups scale O(1).

> - Thus, the routing table really needs to be constrained to
>   grow at about Moore's law for memory.
Not for packet forwarding. The problem is processing the routing 
updates. Since the number of updates scales linearly with the number of 
routes, but you need to look up a route in order to process it and then 
possibly do some data structure manipulation as a result, this scales 
O(n*log(n)) which is not good.

> - If the exceptions are growing at about 100% per year, and
>   the memories are growing at about 100% every TWO years, then
>   regardless of the starting point, the exceptions will overtake
>   technology.
Unless we can make sure that the lines terminate before this happens. 
If we assume 10% of the population will multihome, there is an 
automatic limit around 1G multihomers around 2050. However, one person 
may want to multihome several times by then...

> - Therefore, we must find some mechanism that prevents the
>   exceptions from growing at 100% per year.  In short, the
>   number of longer prefixes that are injected into routing
>   cannot be a constant fraction of the number of sites that
>   join.
Or we introduce another variable to bend the curve downwards. By no 
longer keeping a copy of the full global routing table in every 
individual router, but distributing it over a large number of routers, 
we can make sure we don't run out of memory (or CPU power) in the near 
future.

> - Since everyone and their brother will want an exception
>   for anything that they want to do that is outside of the
>   norm, the norm MUST support almost every possible situation.
>   Multihoming, in particular, must not cause exceptions.
>   Even a constant percentage of multihomers must not cause
>   exceptions.
Agree. But if we can make multihomers connect to their ISPs and make 
ISPs interconnect within regions multihoming no longer causes 
exceptions. We don't even need all multihomers and all ISPs to conform 
to this, just enough to keep us on the good side of Moore's law.

> - For reasons that I've already explained, the economics
>   of links in a geo system cause many sites to be exceptions.
It's very nice to have a cheap long distance links and multihome to 
distant ISPs, but what does that buy you if you can't be routable? 
Economics is about making rational decisions, not about forcing huge 
costs in one area to obtain slight savings in others.

Besides, if links from Jersey City to Palo Alto are so much cheaper 
than across the Hudson, why not multihome to two ISPs in California and 
be geo aggregatable rather than connect to Palo Alto and NYC and break 
aggregation in the process.

> - Therefore, geo addressing leads to a system that will not
>   scale for the long term.
> QED
The real reason geo won't scale in the long run is that at some point 
the number of multihomers in a city becomes too large, and I don't 
believe it is possible to do reasonable geographical aggregation within 
a single city.

But let's not compare apples and oranges. I agree that geo aggregation 
won't solve the long term problem. However, I does offer short term 
relief and intermediate term disaster relief. If we can make every IPv6 
enabled host use multiple addresses for every application within two 
years, that would be much better, but I don't see this happening fast 
enough that we can do without a short term solution.

Iljitsch