Re: how many locators?

[Masataka Ohta <mohta@necom830.hpcl.titech.ac.jp>]

Andrew Partan;

First, thank you for initiating a real engineering discussion,
with quantatative analysis possible.

> How many locators should hosts have

At Vienna, I said 9 (=3*3) is a proper upper limit.

> - or - how big is the global
> routing table

1K~8K.

> - or - where do you get your locators from?

Anyway, though DNS is suggested.

> As the number of locators per host goes up, the complexity of
> choosing a pair of locations to use for a onnection goes up even
> faster (probably as n**2).

No. It is n. We have code of our protocol to do so.

Moreover, global routing table small enough to be held on hosts
can optimize the order of try that n can, in most cases, 1 or 2.

> So we probably want to limit the number of locators per host to a
> smallish number.

Still yes, though 16 or 20 may be tolerated.

> Now hosts end up getting their locators from their upstreams (and
> their upstreams' upstream?).  The question is how far up you go.

Two levels of ISP is reasonable.

But, note that lower level ISPs can peer with a lot more than
3 other ISPs.

A site may directly negotiate with upper level ISPs to have
9 upper links.

Depending on a policy to assign TLAs, big companies may get
its own TLA and be multihomed with IPv4 style. But, it should
cost a lot.

> If you go to the upstream's upstream's upstream, then you are
> approaching approx the Tier-1 core, so the size of the global routing
> table is approx the size of the Tier-1 core -- but at the cost of
> having lots of locators per host.

The cost should be paid by those wanting to have lots of locators
or their peer.

> If you go just one hop to the upstream to get the locator, then you
> have just a few locators per host (basically the number of transit
> connections per site), but you have a much larger global routing
> table - approx the number of ISPs in the world.
>
> Is this a reasonable size with reasonable growth factors?

Currently, ISPs are assigned address blocks, primarily if they
are multihomed to other ISPs, which is inevitable with multihoming
issues unsolved.

As the number and density of ISP grows, more and more ISPs qualify
to have its own address blocks. Note that, with dark fiber, cost
of local peering is almost zero.

Suppose that an ISP, in average, serves 1K people mostly singly
homed and world population is 5G, there will be 5M ISPs.

Do you think it small enough?

						Masataka Ohta