Guohai,
I agree with Richard¹s earlier comment that the consensus of the WG was
that an endpoint is in one PID. Here are my reasons for supporting that.
I believe that clients do not care about PIDs. PIDS are irrelevent.
Clients really care about endpoints. We introduce PIDs because an
endpoint-to-endpoint cost matrix is too large (for today¹s computers,
anyway!). So we partition the 2^32 (or 2^128) addresses into a few
hundred, or maybe a few thousand, equivalence classes, called PIDs, and
have a cost matrix from PID-PID rather than endpoint-endpoint. So when a
client wants to cost from EP1 to EP2, the client maps each EP to a PID,
and gets to costs between those PIDs.
But that makes it essential that an endpoint is in only one PID. If it's
in several PIDs, then that endpoint has several different costs. Which one
should the client use?
It is like the old observation that a person with one watch always knows
the time of day. But a person with two watches is never sure. :-)
So whenever someone suggests an extension that would allow an endpoint to
be in multiple PIDs, my immediate question is, What would a client do with
that information? Why would a client care?
Incidentally, what if the server offered multiple Network Maps, with
different endpoint partitions, and different cost maps? That certainly is
supported by RFC 7285, and (in a sense) allows an endpoint to be in more
than one PID. However, it then raises the question of how the client
decides which Network Map to use.
- Wendy Roome
On 04/14/2015, 15:00, "[email protected]" <[email protected]>
wrote:
>Message: 1
>Date: Tue, 14 Apr 2015 04:50:02 +0800 (GMT+08:00)
>From: ??? <[email protected]>
>To: "Y. Richard Yang" <[email protected]>
>Subject: Re: [alto] ALTO extension for representing SDN policies
>Message-ID: <[email protected]>
>
>Dear Richard,
>
>The longest prefix matching mechanism in ALTO maps each IP address into a
>single PID, this is the non-overlapping property of an ALTO network map.
>But in the end-to-end
>policies such as GBP, EPGs may be overlapped, some endpoints may belong
>to multiple EPGs, the longest prefix matching mechanism should be changed
>to adapt for that.
>I think that one prefix or IP address can be mapped into a PID list that
>contains one or multiple PIDs. We can still use the longest prefix
>matching, but the result is a PID list,
>not a single PID.
_______________________________________________
alto mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/alto
