RE: How do we handle 4 billion CRH-FIB entries ¿?

[Ron Bonica <[e-mail]>]

Pablo,

The scope of the current draft is limited to CRH forward plane behavior and CRH-FIB scalability. Your question concerns the scalability of a path computation module. While your question is legitimate, it is beyond the scope of this document and the charter of this working group.

Nonetheless, I will try to answer it. Section 7 of this document says that although identifiers have node local significance, they can be assigned as if they had domain-wide significance. In my last email to Robert, you will see that I allocated loosely routed identifiers as if they had node local significance.

In many cases, this strategy is beneficial and facilitates scaling. You will see it in the IS-IS extensions for CRH <https://datatracker.ietf.org/doc/draft-bonica-lsr-crh-isis-extensions/> document. When this strategy is used, control planes that support CRH have roughly the same scaling characteristics as control planes that support SR-MPLS.

                                                             Ron






Juniper Business Use Only
From: Pablo Camarillo (pcamaril) <[e-mail]>
Sent: Friday, May 22, 2020 7:00 AM
To: Ron Bonica <[e-mail]>; [e-mail]
Cc: [e-mail]; Ole Troan <[e-mail]>; Robert Raszuk <[e-mail]>
Subject: How do we handle 4 billion CRH-FIB entries ¿?

[External Email. Be cautious of content]

Hi Ron, authors,

I've been reading the example you provided to Robert/Ole, and there's something concerning to me:

> Consider the following network:
>
>      -    Contains 65,000 routers
>      -    Each router has 500 directly connected neighbors or fewer
>      -    Uses 16-bit CRH
>
>In this network, each node might have 65,499 CRH-FIB entries:
>
>     -     64,999 CRH-FIB entries cause packets to follow the least-cost path to another node in the domain
>     -     500 CRH-FIB entries cause packets to traverse a specific link to a specific neighbor.

In addition, you've said several times that:
> Now recall that identifiers have node local significance

In summary:
You have 65000 routers in your network, and each router has 65499 "node-local" CRH-FIB entries.
In other words, this network has in total 4257435000 node-local CRH-FIB entries.

1.- How does it scale for an operator to configure on CLI more than 4 billion node-local CRH-FIB entries for the entire network?
2.- Assuming that some day in the future you would like to advertise this in the IGP: how is it going to deal with 4 billion node-local entries?

Also, you said:
> Somewhere in the network, there is an entity constructs the CRH and the list that it contains. That entity needs access to another data structure, that includes a global view of each node's CRH-FIB.

So, this entity needs a data structure with 4 billion node-local entries, and I guess some path computation algorithms with that scale...

Note that as per your explanation, we could configure the same CRH-FIB entry/tag on several nodes. However this changes nothing: as long as entries are defined as node-local you still have to deal with 4 billion.

In my opinion this is a fatal scalability problem.

Thanks,
Pablo.

PS. Those numbers are only assuming 65.000 routers; however as part of the 16b vs 32b the authors have stated several times the need to support more routers...