Shim6 vs PI addressing

Fri Mar 3 18:50:03 UTC 2006

Thus spake "Iljitsch van Beijnum" <iljitsch at muada.com>
> On 1-mrt-2006, at 18:05, David Barak wrote:
>> Is it easier to scale N routers, or scale 10000*N hosts?
...
> 2 x relatively small is a lot less than 10 x relatively large. Or, in 
> other
> words: on the host you only pay if you actually communicate. In
> routers, you pay more as there is more routing information, whether
> the extra information is used or not.

OTOH, hosts go a lot longer between upgrades and generally don't have 
professional admins.  It'll be a long, long time (if ever) until shim6 is 
deployed widely enough for folks to literally bet their company on 
host-based multihoming.

>> If we simply moved to an "everyone with an ASN
>> gets a /32" model, we'd have about 30,000 /32s.  It
>> would be a really long time before we had as many
>> routes in the table as we do today, let alone the
>> umpteen-bazillion routes which scare everyone so
>> badly.
>
> 1. We've already walked the edge of the cliff several times (CIDR had  to 
> be implemented in a big hurry, later flap dampening and prefix  length 
> filtering were needed)

At least this time we know what the likely problems are, and we can build in 
safeguards that can be quickly implemented if we get too close to the edge. 
Not that I agree we'll even get there...

> 2. We'll have to live with IPv6 a long time

Perhaps.  I know the goal was for it to last 100 years, but what technology 
has ever lasted that long without significant improvements that altered it 
almost beyond recognition?

> 3. Route processing and FIB lookups scale worse than linear

With an mtrie+ FIB, routing lookups scale far, far better than linear.  What 
matters is the length of the prefix being matched, not how many there are.

TCAMs scale linearly, provided you can build them big enough (and costs 
certainly aren't linear).

> 4. If the global routing table meltdown happens, it will be extremely 
> costly in a short time
> 5. Even if the meltdown doesn't happen a smaller routing table makes 
> everything cheaper and gives us more implementation options (5000  entry 
> TCAM is nice, 500000 entries not so much as it basically uses  100 times 
> as much power)

Agreed.

> 6. Moore can't go on forever, there are physical limitations

Every time folks claim that, someone makes a breakthrough that continues the 
curve.  Surely we can't count on this forever, but so far money has 
consistently trumped "physical limitations".

> But the most important thing we should remember is that currently, 
> routing table growth is artificially limited by relatively strict 
> requirements for getting a /24 or larger. With IPv6 this goes away,  and 
> we don't know how many people will want to multihome then.

The requirements for getting a /24 are pretty darn lax, actually, and the 
current proposals for PI space being debated within ARIN are significantly 
more restrictive.

The reality today is that v4 routing tables are well within our capabilities 
and growing slowly.  If we were on the verge of another serious problem, 
like we where when the CIDR fire drill happened, ISPs could easily cut the 
tables in half simply by filtering prefixes longer than RIR minima.

S

Stephen Sprunk        "Stupid people surround themselves with smart
CCIE #3723           people.  Smart people surround themselves with
K5SSS         smart people who disagree with them."  --Aaron Sorkin 