At 11:38 14/02/2005 -0500, Alia Atlas wrote:
At 06:33 AM 2/14/2005, mike shand wrote:
At 21:14 13/02/2005 -0500, Alia Atlas wrote:
The LFA draft currently doesn't really discuss ECMP and the protection
that may or may not be available via ECMP. For a particular destination
D, S may have multiple equal-cost primary next-hops.
In the following figure, S has three equal-cost paths via E1, E2 and E3.
[N]-----[ S ]--------[E3]
| 5 | |
20 | | |
| --------- | 2
| 5 | | 5 |
| [E1] [E2]------|
| | |
| 10 | 10 |
2 |--[D]--| 2
In this example, the primary next-hop to E1 can get link and node
protection from the primary next-hop to E3; it can only get node
protection from E2. E2 can't get node protection from E1 or E3; E2 can
get node protection and link protection from the loop-free neighbor N.
Simply having multiple primary next-hops does not guarantee the desired
type of protection. The basic question is whether it is preferable to
use an equal-cost path that doesn't provide all the desired types of
protection or an alternate path that does give the desired protection?
Should the draft discuss this issue or describe the use of other primary
next-hops as alternates in any more detail?
Yes, I think so. We had some text in our tunnels draft which addressed
some of this (section 4.6) , but the LAN case as you describe makes it
more complicated. I think if we are going to say you can use ECMP at all
(which I think we should), we need to say exactly under what conditions
it is OK.
So describe the above example as a case of how ECMP doesn't guarantee
protection - and then up to the implementation/operator to decide what the
preference between protection type and cost is?
Don't we already have to discuss the same sort of issues with respect to
LFAs? ie. you may be able to get a viable LFA where you couldn't otherwise
by relaxing the protection type.
But if it were a choice between using an ECMP which may not actually
protect against the failure, and using an LFA which is guaranteed to
protect against any failure, then I think I would probably choose the
latter. It is only in cases where there was no "perfect" LFA that I might
be tempted to use a less than unconditionally OK ECMP (or even LFA).
But as you say, to some extent it is an operator/implementation choice.
Rtgwg mailing list
Rtgwg mailing list