Re: [DMM] Comments on draft-patil-dmm-issues-and-approaches2dmm-00

[jouni korhonen <[e-mail]>]

Pete,

On Mar 7, 2012, at 4:55 PM, Peter McCann wrote:

> Hi, Jouni,
> 
> jouni korhonen wrote:
>> Pete,
>> 
>> Thanks for the review. Some thoughts inline.
>> 
>> On Mar 6, 2012, at 11:48 PM, Peter McCann wrote:
>> 
>>> Hi, Raj, Carl, and Jouni,
>>> 
>>> I have some comments on draft-patil-dmm-issues-and-approaches2dmm-00.
>>> 
>>> I agree with most of Section 4, "Issues with current mobility models". 
>>> However, I'd like to point out that existing networks are not just
>>> centralized in the manner you point out, they also tend to have a
>>> hierarchical structure, e.g., the S-GW/P-GW split in 3GPP EPC. 
>>> Therefore, the issue you outline in
>> Section 4.3
>> 
>> It is quite common to run combined nodes.
> 
> Sure, in that case the combined S-GW/P-GW is a centralized anchor
> point and the excess signaling you point out in Section 4.3 would
> indeed be a problem.

Actually, signaling due a handover (SGW/SGSN relocation or even L2 handover)
in PGW/GGSN is not the biggest issue.. It is the bearer management in general,
which is not a problem for IETF to tackle. However, designing a system that is
conservative on signaling is a good general guideline. And that is the reason
we emphasize that. So not let us get too stuck with EPC, rather learn from it.


>>> ("Inefficient Routing and signaling overhead") is not quite true of the
>>> 3GPP EPC, which can handle many mobility events in a localized manner
>>> similar to HMIP.
>> 
>> Could you clarify which functionality in EPC you refer to from IP point
>> of view?
> 
> I mean the ability to update a local S-GW with each eNB change, which
> avoids the extra signaling to the P-GW.  It is architecturally similar
> to HMIP (really, PMIP + HMIP).

Comparing against HMIP is not that straight forward. In HMIP you have
IP exit points at AR, MAP and HA, also local mobility at the IP level
under MAP and L2 mobility under AR. With EPC, you have one exist point
at your gateway and L2 mobility under SGW.

>>> The first paragraph of Section 7 talks about source address selection,
>>> and the need to modify applications so that they request the kind of
>>> address that they want.  I tend to think that applications will remain
>>> unmodified for some time to come; however, most applications fit into
>>> the paradigm of opening short-lived connections to a server and could
>>> be accommodated with some sort of automatic handling in the MN's IP
>>> stack.
>> 
>> I tend to think that developers who care and see some benefit for their
>> applications would update. Completely automated solution within the
>> stack would be nice but that also entails larger MN update and would
>> also need some additional (out of band) policy information to guide
>> address selection.
> 
> It should certainly be possible for both models to co-exist.  If an application
> specifies an address, that's great, but if it doesn't, maybe it should just
> get a fixed privately scoped address that is then NATted out to whatever
> public address the MN happens to have at the time the connection gets made.
> 
>>> I found the last paragraph of Section 7 quite interesting.  I too think
>>> that there is an important piece missing that you call "seamless
>>> mobility anchor relocation". I think that the use of an interior
>>> routing protocol is spot on.  In fact, if you read
>>> draft-mccann-dmm-flatarch-00, I propose just that.  I think we can use
>>> such an anchor relocation protocol to make each access router in an
>>> autonomous system (or smaller region of an autonomous system) a
>>> temporary anchor for a given prefix. In my draft I propose running
>>> I-BGP on each AR and sending BGP UPDATES into the network upon
>>> localized mobility events.  Such a protocol can also be
>> used as a
>> 
>> I yet need to read your draft properly.. but in general trying to solve
>> mobility issues with routing protocols is just moving the problems out
>> of your hands to others. You easily end up with uncomfortable amount of
>> host routes and your network being constantly in a "converging" state..
> 
> We need to do some simulation/implementation experiments to see if this
> is a problem.  Hopefully I'll have more to say on that soon.  But, intuitively,

That would be excellent.

> the UPDATE just has to propagate one hop to the crossover router to be effective.
> We don't need the whole network to be converged for packets to take the right
> path toward the currently serving AR.

You need to be careful on the deployment architecture you introduce
stuff like this. I am just being conservative ;) Once the constraints
and assumptions are better known it is easier to evaluate the solution.

>>> substitute for the proxy ND technique that is currently specified to
>>> "grab" the MN's packets at the HA.  By using a routing protocol, the HA
>>> can reach across several routing hops so it doesn't necessarily need to
>>> be on the home link (which can be the first AR to which the MN
>>> attached).  I think this would also enable us to unify the
>>> authentication protocols used at the AR with the authentication
>>> protocol used at the HA.  The ARs are just like HAs that don't have to
>>> tunnel the data anywhere because the MN is locally connected.
>> 
>> There are virtues for exploring the routing protocol "enslavement" for
>> additional mobility management. Geo-redundancy would, for example,
>> benefit from it.
> 
> You would get all the benefits of distribution and fault-tolerance from
> the routing protocol itself.  IP routing already has these features built-in.
> It is the establishment of tunnel state in a single box that makes the network
> fragile.


- Jouni



> 
> -Pete
>