Need for N-homedness (was Re: asymmetric routing)

[Pim van Pelt]

| Well, I am not 100% sure how the big guys solve this (that is,
| probably most of you), but what about redundancy in case of
| malfunctions? Say, someone digs through a cable (don't laugh! This has
| happened to me!). Or some other failure that you have no way to
| control.
| 
| Such problems can be completely unrelated to whether you are using
| IPv4 or IPv6. In fact, I haven't ever heard of anyone cutting through
| a cable by accident first asking "OK, what transport protocol is being
| used on this one?"
| 
Michael, others,

I believe it is a common misconception that having redundant uplinks have
something to do with which protocol you speak and, having redundant uplinks
(two fiber pairs to your ISP(s)) also has nothing to do with dualhomedness.

Riverstone for example, has a 'smart trunk' functionality, which means 
combining two or more physical ports (ether, pos, atm, gige) into one layer3
endpoint, enabling packet-for-packet loadbalancing over two lines, or some
crude (at that time) form of QoS routing, by sending realtime over link A and
lower prio over link B. We used to have two E1 links between two locations,
with the restriction that both links started and ended in the same physical
box (note the SPoF here).

You can easily have multiple fibers with layer2 also, think of fast spanning
tree or other (faster) protocols like VRRP. Again, no need for more than one
aggregating ISP.

The ISP can buy transit from more than one carrier, can set up iBGP sessions
with two of the customer (downstream) routers via two cables, for all I care:

TRANSIT1-(R1)------fiber-to-customer-----(R3)---\
          |                               |     |------network at customer
TRANSIT2-(R2)-----backup-E1-to-customer--(R4)---/
(fear my ascii art - best viewed in xterm, not in Outlook et al :)

ISP has R1/R2, interconnected and each one having an uplink with a full view
of the dfz via seperate transit providers. Customer has R3/R4, each linking 
to an ISP via a seperate physical circuit, and handling the customers network
via VRRP (a protocol which switches 'the default gw' between R3 and R4 if 
either one fails). Most new hardware does this.

I fail to see any need for more than one prefix per customer, and still say
that IF a customer has more than one prefix, his R3/R4 should preform source
based policy routing, sending traffic from prefixA through the circuits to
ISPA and traffic from prefixB to ISPB. 

groet,
Pim

-- 
---------- - -    - - -+- - -    - - ----------
Pim van Pelt                 Email: pim@ipng.nl
http://www.ipng.nl/             IPv6 Deployment
-----------------------------------------------