RE: GSE

["Tony Li" <Tony.Li@procket.com>]

Iljitsch,

|   > Each host needs to advertise its locator to correspondent 
|   hosts.  Whether
|   > this happens via DNS or via a mobile IP type solution is 
|   more detailed
|   > than we need to be right now, but basically, the correspondent can
|   > get multiple locators for any MH host.  Failover in this 
|   case is simply
|   > changing the locator in the packet and does not require tunneling.
|   
|   Yes, switching to a different locator is what we want to do when a
|   failure occurs. However, the GSE draft doesn't.


Fine, as I said, I'm looking at what we would like to do, not just
what GSE currently does.  Again, please see previous mail on 16+16.

   
|   Determining when to change locators isn't an altogether trivial
|   exercise. This is where we can use the help from TCP, since it knows
|   about timeouts and retransmissions, the IP layer has no 
|   idea about any
|   of this stuff.


Yes, exactly.


|   > |   - The flat part of the identifier space makes locator
|   > |   discovery hard.
|   
|   > Identifiers need not be flat.  They could be 
|   hierarchically assigned
|   > administrative entities.
|   
|   EUI-64 identifiers don't have a hierarchy that is of use to 
|   us. Using
|   anything else means autoconfiguration has to be changed.


There are enough issues around EUI-64 that I thought it was
good to get away from those anyway.  

In any case, I'm not convinced that there's a LOT of need to
have an identifier to locator mapping.  You need a hostname to 
identifier and locator mapping, but when would you go from 
identifier to locator without the hostname?

In any case, if we were to really want this, yes, autoconfiguration
would have to change.  The identifier space could have hierarchy,
but need not be tied to the topology.  Further, it wouldn't need
to aggregate except to make entry into DNS palatable.  The
autoconfiguration would assign an identifier upon recognizing the
host based on the range of identifiers available to that particular
subnet.


|   None of this is impossible, but what exactly is it about 
|   the 6+2+8 thing
|   that makes it worth going through all this trouble? If you 
|   simply assign
|   people a PI /48 you can use this address block as identifiers and
|   replace the first six bytes to make locators. The only 
|   "problem" with
|   this is that you have to replace the first six bytes in the 
|   address with
|   the original ones at some point, but it's much easier to make that
|   happen than having transport protocols only look at the 
|   bottom 64 bits.
|   If only because some box at the edge can do this.


Why is it hard to make the transport protocols only look at the 
bottom N bits?  Ok, yes, you have to feed them into the
checksum algorithm separately, but this is NOT rocket science.

Tony