Re: [pilc] IESG Review of LINK

[Alex Zinin <zinin@psg.com>]

Joe,

>>  > 17 Routing
>>  > 
>>  > Many subnetworks provide their own internal routing mechanisms.
>> 
>>  The text should probably say why some do and some don't. E.g., if a
>>  subnetwork technology can be used to construct a multi-node network,
>>  where internal nodes do not run IP, and where the network is
>>  self-contained and can be used for service independent from IP
>>  routing, then subnetwork-level routing function is usually provided
>>  (FR, ATM, Ethernet bridging and STP). If the technology is primarily
>>  used for single-hop packet framing among IP nodes, subnetwork-level
>>  routing is not needed.

> The issue isn't whether it's a multi-node network; the issue is whether 
> the media is switched or not.

Switching is a consequence of the decision to provide the routing
function within a network, and one doesn't need routing for a
single-link technology... Anyways, this is not a terribly important
argument here.

> Possible text: (?)
> --------------------------------
> Multipoint subnetworks may be single-hop or multi-hop. In either case, 
> IP considers the subnetwork as a single shared link, i.e., IP assumes 
> there is one or more subnetwork addresses that can broadcast to all 
> nodes on the subnetwork with one subnetwork packet. IP also assumes that 
> multi-hop multipoint subnetworks provide their own routing, such that IP 
> can inject the subnetwork packet with the appropriate destination 
> subnetwork address, with no further path information.
> --------------------------------

The "single link" and "broadcast" assumption above are not always
correct. One could model connectivity through a cloud as a set of p2p,
p2mp, and broadcast-emulation logical segments. Also, in many
subnetwork technologies there no broadcast capabilities (so-called
NBMA.)

>>  > Since routing is the major function of the Internet layer, the
>>  > question naturally arises as to the proper division of function
>>  > between routing at the Internet layer and routing in the subnet.
>> 
>>  and the text does not address this question, nor does it talk
>>  about the interactions among the routing function within the
>>  subnetwork and at the IP level.

> I believe the above addresses the interaction - i.e., that in general IP 
> assumes there is NO interaction, that the subnetwork can be considered a 
> single shared link and be done.

I don't believe this is correct.

>>  > In general, routing in a subnetwork and at IP is more 
>>  > complementary than competitive. Routing algorithms often have 
>>  > difficulty scaling to very large networks, and a division of labor 
>>  > between IP and a large subnetwork can often make the routing 
>>  > problem more tractable for both.
>> 
>>  In fact, experience shows that this is exactly the opposite.
>> 
>>  When a large number of IP routers is connected via a subnetwork that
>>  hides internal routing details from IP, this usually results in O(N^2)
>>  adjacencies, flooding storms, and unforeseen convergence interactions.
>>  In most cases, the physical topology of a subnetwork is orders of
>>  magnitude simpler than the logical IP topology on top of it. This is
>>  because the topology graph within the subnetwork is determined by
>>  physical links and nodes, while the topology graph at the IP level is
>>  essentially driven by the connectivity matrix.

> But the link connectivity matrix is insufficient to express the 
> reachability of IP addresses, which is what IP needs to compute.

I meant connectivity matrix of the subnetwork.

My point was that "division of labor" mentioned in the text, does
NOT make the routing problem "more tractable". In fact, it makes
it worse.

>>  With the same number of nodes, routing scales better if it operates
>>  over the physical topology rather than when it's layered up.

> Link routing doesn't scale when viewed as a transit; in that case, the 
> link subnet looks more like an AS, and IP routing convergence faulters 
> when interior routing protocols (e.g., ones that assume broadcast or 
> multicast) are used over a multihop subnetwork.

> These are good reasons not to use large-scale subnetworks as transits in 
> dynamically routed IP networks, or to better manage the route 
> advertisements in a BGP-like fashion.

Frankly, didn't have a clear feeling whether you agree or disagree
with me.

>>  OK, meta issues now.
>> 
>> What I would expect this section to talk about is:
>> 
>> 1. Difference between the _function_ of IP routing and the routing
>>       function found in some subnetworks and analysis of why
>>       certain subnetwork technologies have the routing function.

> (see contributed paragraph above)

>> 2. Give a recommendation that unless there is a very compelling
>>       reason, routing should be left to IP.

> (see contributed paragraph above)

didn't see text trying to address this

>> 3. If the routing function is still implemented in the subnetwork
>>       technology, explain what considerations should be kept in mind:
>> 
>>    a) Connectivity graph and its affects on IP routing, such as:
>>            - number of adjacencies/sessions per router to maintain
>>            - broadcast & connectivity (if routing treats the subnet as
>>                broadcast, any-to-any connectivity needs to be ensured)
>>            - flooding implications
>>            - ...

> Perhaps a paragraph warning about how transit subnetworks end up looking 
> like AS's ;-)??

hmmmm... I think not... It would be better for the doc to discuss
IGPs, as we normally have IGPs running on top of a subnetwork.

>>    b) Rerouting, convergence and timing with multiple layers:
>>            - effects of topology change and rerouting within subnetwork
>>                on IP routing (e.g., if the subnetwork has slow rerouting,
>>                IP will start reconverging)
>>                    
>>            - effects of reconvergence at the IP level on the subnet
>>                (e.g. dynamic connection establishment)
>> 
>>            - combination of the two

> This hits a different concept. IP doesn't really like having links 
> appear and disappear; it expects the subnetwork to be largely static. 
> That point might be useful to raise (?).

Right, from the IGP perspective, a change in topology is a transient
event and stable topology is the normal condition. Note however,
that the issues above are related to the interaction of the routing
functions.

>>    c) Connection model for connection-oriented subnetworks
>>        
>>          If the model assumes temporary connectivity between
>>          nodes (e.g., PSTN or ISDN) how will this affect IP routing
>>          (adjacencies going up/down, applicability of e.g. demand circuit
>>          extension in OSPF)

> Such systems can be 'faked' by having routing proxies on both sides of 
> the idle links that lie about reachability, and connect only when there 
> is traffic in place. I.e., when the links are made on-demand, it can be 
> useful to let IP consider the links as always there, and kill the 
> proxies only when connection establishment fails.

Basically, yes. I would recommend studying the approaches that RIP and
OSPF took for Demand Circuit extensions. The fundamental concept is
presumption of reachability and fake adjacencies.

>>    d) Optimality and traffic engineering
>>        
>>            it's addressed in the text to some extent, but more thought
>>            should be given to such aspects as how traffic engineering
>>            within the network (pure IP and MPLS) would be done, i.e.,
>>            is it just "enough BW", or the subnetwork provides TE
>>            functionality as in ATM, for example...

> Shifting sands again. ;-)

:) Quite concrete, actually. The question is "how do I manage
available BW of the subnetwork?"

Alex


_______________________________________________
pilc mailing list
pilc@ietf.org
https://www1.ietf.org/mailman/listinfo/pilc
http://www.ietf.org/html.charters/pilc-charter.html
http://pilc.grc.nasa.gov/