[Ospf-wireless-design] Response to comments on draft-ogier-manet-ospf-extension-03

[Richard Ogier <[e-mail]>]

All,

Tom summarized some concerns about my draft that were voiced at the
OSPF session on Tuesday, which I would like to respond to.
The main concern seems to be that the solution may not be robust
because there are not enough Backup DRs, or because each DR has too
many Dependent Neighbors, or because it may take a long
time for database exchange when a new DR is selected.
I don't think these concerns are valid, but we will find out
with simulations (and I agree with Joe that simulations should
include failure models in addition to mobility). It would also help
if people with these concerns provide specific examples on the
ospf-wireless-design list; that way, we can be sure the concern
is real, and we will have concrete examples that we can discuss.

For now, I will clarify some things that might alleviate any
concerns about robustness.

1. There might be a misunderstanding that data packets can only
   be routed on adjacent (or fully ajacent) edges. This is not
   true in my draft, nor is it true in OSPF (in a broadcast
   network, a non-DR/BDR router can forward a data packet to
   a non-adjacent neighbor, as long as both routers are adjacent
   to the DR - note that the "synchronized neighbor" condition
   in my draft generalizes this condition).
   For example, in the full topology case, LSAs describe the
   full topology even though there are only about 2n adjacencies.

2. A new DR does not have to complete database exchange with its
   neighbors before it can flood LSAs. (This is also true in
   legacy OSPF).

3. If a DR fails, then a BDR will forward the LSA for it, so
   the main concern is what happens if a DR and a nearby DR/BDR fail
   at nearly the same time. The probability of this happening is
   fairly low, but it can happen, so let's see what happens.
   Note that it is possible to select more BDRs for more redundancy,
   or to allow *any* router act as a BDR, but I don't think this
   will be necessary, for the following reason.
   With differential Hellos being sent every 1 or 2 seconds (which
   is quite possible since they are small), if a DR fails, then
   a new DR will select itself within a few Hello intervals, e.g.
   between 3 and 6 seconds.  Since RxmtInterval is typically 5
   seconds, the new DR will probably select itself before the
   LSA would be retransmitted anyway!

I would also like to answer a question regarding what happens
if the 2-hop neighbor information is not consistent.
This is a good question, and I will update my draft to clarify
this. If node i has two neighbors j and k such that j indicates
(in its Hellos) that k is a neighbor, but k indicates that j is
NOT a neighbor, then node i must assume that j and k are NOT
neighbors (i.e., both neighbors must agree that they are
neighbors of each other). This provides faster recovery from
a link failure, since one neighbor may detect the failure
sooner than the other neighbor.  (Also, note that the rule
is different for partial 2-hop neighbor information, since
the absence of a neighbor RID in a partial Hello does not
indicate that the router is not a neighbor.)

Again, if you still have concerns about the CDS solution, please
give specific examples so that we have something concrete to discuss.

Regards,
Richard


_______________________________________________
Ospf-wireless-design mailing list
[e-mail]
https://www1.ietf.org/mailman/listinfo/ospf-wireless-design