On Fri, Mar 16, 2018 at 12:17 PM, Dino Farinacci <[email protected]> wrote:
>> Such complexity is why I am still keen on the redirect model for a
>
> I hear you loud and clear. But we do the redirect model in LISP in many forms 
> as well.
>
>> mapping system. An ILA cache is an optional element and the control
>> plane is never inline with packet forwarding and packets are not
>> dropped on a cache miss. Neither does the generate request packets for
>
> We did that in the ITR as well. A cache missed meant to send a Map-Request 
> and to encapsulate the packet to a PETR (proxy decapsulator) where the PETR 
> usually had a full cache (how it got populated could be with pull or push 
> mechanisms).
>
> But this results in duplicate packets going to the destination as well as out 
> of order packets.
>
>> bogus addresses that can't resolved. These properties bound the worse
>> case DOS attack to be that legitimate traffic takes an unoptimized
>> route but is not blocked nor dropped. Conservatively, this does
>
> Yes, understand. But even in your constrained “domain”, there may be just too 
> much state to push to all nodes. Especially in the 5G use-case. It wasn’t a 
> problem in the LISP beta network because the proxy xTRs had relatively coarse 
> prefixes that reached lots of EIDs.
>
The state would need to be sharded. You'd probably need to do this
anyway for mapping-servers or high thoughput Internet facing routers
for which using a cache would be challenging.

Tom

>> require provisioning ILA-Rs to handle the full load if necessary to be
>> robust.
>
> Yes indeed.
>
> Dino
>

_______________________________________________
lisp mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/lisp

Reply via email to