[nbs] Synthesized names as session-labels - "Was: Re: I-D Action:draft-ubillos-name-based-sockets-03.txt"

[Javier Ubillos <[e-mail]>]

Rémi and Jan.

About the whole "synthesized ip" issue.

I have thought about it a lot and I'm curious as what you think about
the methods along the following lines of thought.

The reason I like the synthesized addresses at all is that it's easy to
derive from something innate to the host. You don't need any
pre-negotiation and it _might_ be better than a randomly generated
nonce.

However, two issues have made me rethink a bit:
* When working with TCP, two names on the same host cannot share IP:Port
pairs. The remote host, if it is not upgraded, wouldn't be able to
distinguish the two flows. There for, a host needs some book-keeping
mechanism to keep the 4-tuples unique.
* Locally, we need the session-identifier to be unique as well to be
able to keep the separate sessions distinct, but also manageable so that
sessions who wish to multi-home/multi-path are able to find and re-use
the correct session label.

So, in other words:
We need to be able to keep a record of which sessions use which
srcIP:dstIP:srcIP:dstIP pairs (or whichever identifier a particular
protocol uses)

What needs to be solved is:
* How to name unnamed nodes
* How to generate sensible session-labels
   * To be able to separate sessions
   * And still allow to aggregate "sub-sessions" (e.g. a multi-path
sessions)

Personally, I think "sub-sessions" are a non-issue, as in the multi-path
TCP case, we can still run standard TCP-sockets as each sub-socket to
the master-socket which will need a session-label.

A thought I'm playing with is to use a combination of a synthesized
names and SRV-record syntax. 
This is an incomplete solution, let me show by an example:

Host A communicates with Host B.
Host A has IP(A) and wants to talk TCP/HTTP to B.
So, it generates 

src name: _<sourceport>._tcp.<IP(A)>.ipv6.arpa
dst name: _http._tcp.<IP(B)>.ipv6.arpa
src IP: IP(A)
dst IP: IP(B)
src Port: _<sourceport>
dst Port: 80

This way, when spawning new sockets, one can by just looking at the
local name, avoid ip:port conflicts.

However, this breaks in the mobility case, where this generated name no
longer has a relation to the ip:port pairs used.

Do any of you have any good suggestions on how to achieve the goal?
Perhaps an overlap in session-labels is not a problem at all (maybe we
can separate sockets at an application level, where no two applications
may interfere with each others sockets...

Do you guys have any good ideas/suggestions?

// Javier




On Tue, 2010-10-19 at 16:15 +0200, Javier Ubillos wrote:
> On Tue, 2010-10-19 at 16:03 +0200, Rémi Després wrote:
> > Le 19 oct. 2010 à 15:03, Javier Ubillos a écrit :
> > >>> ...
> > >>>   o  ip6.arpa.  Using one of the hosts interfaces addresses as a name.
> > >> 
> > >> Why would I want to use this?
> > > 
> > > Another thing I should be more explicit about.
> > > This is a construct for when a host does not have a name associated with
> > > it. Then a name can be synthesized for it.
> > 
> > Since translations from addresses to names are generally time consuming, and not necessarily as up to date as those from names to addresses, this should, as  minimum, be an option that applications would have to request.
> 
> The whole reverse-resolution/ip6.arpa infrastructure is not very broadly
> deployed.
> 
> The way I see it, it's comparable with a nonce, but which _might_
> contain some valuable extra information (the IP the communication was
> started with).
> 
> Also, i6.arpa -> name could very well generate multiple answers. How do
> you know which one is right.
> 
> In other words, this is not a huge deal, the intention is to _maybe_
> provide something that _might_ be more informative than an arbitrary
> label.
> If too many finds that it is just confusing and an eye-sore, we'll have
> to reconsider that possibility.
> 
> // Javier
> 
>