On Fri, May 28, 2010 at 12:33 AM, Fred Baker <[email protected]> wrote:
> There can be at the transport or application layers, in two forms. In both 
> cases, the issue is that a network layer address is meaningful at the network 
> layer and not really meaningful above it, but we have collectively not done 
> well in remembering that. For a good rant on the topic, I would suggest 
> reading the works of Saltzer such as
>
> http://www.ietf.org/rfc/rfc1498.txt
> 1498 On the Naming and Binding of Network Destinations. J. Saltzer.
>     August 1993. (Format: TXT=24698 bytes) (Status: INFORMATIONAL)
>
> or listening to John Day.
>
> At the Transport Layer, if the transport considers a session to be with a 
> remote destination at a locator, and the locator changes during a session, 
> the transport layer will not recognize it and as a result the session will 
> fail. That, specifically, is why ILNP specifies that the locator is not 
> included in transport layer checksums; some would argue that it should not be 
> communicated to the transport at all.

Hi, Fred,

Of course, I know at least this much. This is the rationale for all
those ID/Loc separation. And ILNP is one of them.

I'm not here to argue against ID/Loc separation. As far as I take for
granted that ILNP will be our choice by RRG, I'm not arguing against
ID/Loc separation.

As far as I know, some ID/Loc separation was started to improve
mobility; HIP would be one such example.

Other LIS proposal were more focused to multi-homing; Shim6 might be,
if I'm not wrong, one such example.

And here, we're also primarily concerned with multi-homing to improve
routing scalability, hence ILNP.

Now I want to know more in depth how practically ILNP improves on
routing scalability in DFZ, and how significantly?

I was trying to catch this aspect by comparing

   - how ILNP deals with multiple Locators and
   - how the current Internet deals with multiple PA addresses.
>
> At the Application Layer, any use of a network layer address has pitfalls. 
> You might want to review
>
> http://www.ietf.org/rfc/rfc2993.txt
> 2993 Architectural Implications of NAT. T. Hain. November 2000.
>     (Format: TXT=74136 bytes) (Status: INFORMATIONAL)
>
> for a reminder of why IPv4/IPv4 Network Address Translation has been a 
> headache. These issues remain in ILNP when the application uses a network 
> layer address indiscriminately; routing protocols can have difficulties 
> across the boundary between addressing domains, HTTP Referrals may not be 
> meaningful, SIP/SDP can fail when it communicates the address of a 
> communicant, and so on. The rule has to be that IF ONE IS GOING TO USE A 
> NETWORK LAYER ADDRESS one uses the address appropriate to the peer (an 
> "inside" address if the peer is in the same domain, and an "outside" address 
> in other cases), or (much better) applications confine themselves to using 
> application layer identifiers such as DNS names.
>
>> To me, it looks quite equivalent. My host will know of the two in
>> existence, and will switch between the two as need arises.
>
> There I disagree, or disagree in part. My DNS system will need to know of all 
> of the addresses; my host needs to know only if it is responsible for 
> populating them into DNS.
>
> Let me draw a picture for you.
>                                             ,-.
>                         ,---------.        /   \
>                      ,-'           `+-+  +-+    \     ,-.
>           ,-----. +---+    ISP 1    |R+--+R|     :   /   \
>        ,-'       `+DMZ+.           ,+-+  +-+     :  /     \
>       /+-+        +-+-+ `---------'     ;       +---+ +-+  :
>      / |A|           \                  |       |DMZ| |C|  |
>     ;  +-+            :                 | ISP 3 +-+-+ +-+  ;
>     |           +-+   |                 |         | \Your /
>     :           |B| +---+---------.     :         ;  \Domain
>      \ MyDomain +-+ |DMZ|          +-+   +-+     ;    `-'
>       \             +---+  ISP 2   |R+---+R|     ;
>        `-.       ,-'  `.           +-+   +-+    /
>           `-----'       `---------'        \   /
>                                             `-'
>
> In my domain, I have two hosts: my own is A, and B is another one. In your 
> domain, you have some host C. When B or C want to communicate with A, they 
> ask DNS for A's set of addresses. Ideally, B gets A's "inside" address and C 
> gets A's two "outside" addresses; more probably, each will get all three, and 
> sort among them by discovering, in B's case, that A has an address that is 
> "more similar" (matches a longer subset of its prefix) than the other two, or 
> some of the addresses simply don't work. The question is how the DNS system 
> gets A's 3 addresses. A is actually only aware of the "inside" address and 
> will only use that; however, DNS knows the translation to the external set of 
> addresses, and if A is posting addresses via Dynamic DNS, will do that 
> translation when addresses are posted. If the DNS system isn't doing that, 
> then of course A will have be aware of its addresses.

I think you're describing here a case of NAT, e.g., private address.

In my question to Tony, I was specifically confining the case to PA
addresses; multi-home sites in both ILNP and my example of the current
Internet would use PA Locators or addresses.

I'd like to know the performing difference in these two scenarios.

>
>> I assume we're talking about multi-homing, not mobility here
>> specifically in this context. If it's about mobility, I can see that
>> ILNP has an advantage for TCP connection is not associated with
>> locator while it would be with the current IP address.
>
> Actually, multihoming becomes a special case of mobility.

I know, but not completely.

Multi-homing can be considered as slow mobility. With fast mobility,
you'd be sensitive about continuity of TCP connections. With
multi-homing, you'd be more relaxed and only concerned about roaming
aspects. A host in a site-multi-homed host would have multiple PA
addresses, and the host or remote correspondent nodes would be more
concerned about which address to use in a newly initiating connection.

>
>> But as long as you were talking about multi-homing, you would not mean
>> fast transitioning without breaking a TCP connection. You would
>> implicitly mean a situation where the same host would be associated
>> with a connection at one point in time and with another at a different
>> point in time. Or even with two simultaneous yet separate connections.
>
> If the locator is irrelevant at transport and above, this also helps with the 
> multipath TCP case. It requires the application to send not to a given remote 
> address but to all of them, but it has that possibility.

Of course. But, again, let's please focus on behavior in
transitioning/failover across multiple-PA Locators/addresses.

>
>> In all these connection scenarios, I would assume that the effect
>> would be the same in both cases of ILNP and the current Internet.
>>
>> I must be missing a critical point here.
>>
>> But, if I'm not, wouldn't it the description of yours apply in the
>> same way to the current Internet? The same situation with multiple PA
>> addresses?
>
> yes and no; TCP in the present internet cares a great deal about the locator, 
> and applications do as well. Shim6 tries to address those issues, but a shim6 
> internet is not the same as the present one either.

I think, we're shooting at different targets here. Your statement here
is of course correct. My focus is how to reduce the DFZ routing table
size.

-- 
DY
_______________________________________________
rrg mailing list
[email protected]
http://www.irtf.org/mailman/listinfo/rrg

Reply via email to