Re: [mpls] mpls-in-udp entropy

Curtis, 
IMHO and FWIW it is preferable to allocate the entropy port from the Dynamic/Private space.
14K values should suffice for any reasonable ECMP scenarios.

My 2c,
     Sasha
________________________________________
From: Curtis Villamizar <curtis@ipv6.occnc.com>
Sent: Wednesday, January 15, 2014 10:56 PM
To: Alexander Vainshtein
Cc: erosen@cisco.com; mpls@ietf.org
Subject: Re: [mpls] mpls-in-udp entropy

In message <5b0765246d204750a50e1aad52a3b72e@AM3PR03MB532.eurprd03.prod.outlook.com>
Alexander Vainshtein writes:

> Eric,
> Lots of thanks for a prompt and highly informative response.
>
> I have been actually thinking about the same thing, namely that the entropy port should be the result of some hash over the label stack.
>
> If this is indeed the intention of the authors, it would make sense (at least, from my point of view) of saying so in the draft. There is no need to make such a statement normative, but it would really help the readers (both implementors and operators) to understand what it is about.
>
> Regards,
>      Sasha

Avoiding the lower 8K of the port number space might not be a bad idea
to avoid a return port being a WKP including the non-root WKP space
used by X-Windows and other things.

Curtis

> ________________________________________
> From: Eric Rosen <erosen@cisco.com>
> Sent: Wednesday, January 15, 2014 6:35 PM
> To: Alexander Vainshtein
> Cc: mpls@ietf.org
> Subject: Re: mpls-in-udp entropy
>
> (Changed subject line and trimmed cc-list.)
>
> Sasha> I would like to understand whether this protocol can really result in
> Sasha> reasonable distribution of traffic. "Reasonable" means that (a) there
> Sasha> is sufficient entropy and (b) that the order in specific micro-flows
> Sasha> is preserved.
>
> I thought the intention was that the encapsulator would set the UDP source
> port based upon the entropy of the packet being encapsulated.  This only
> requires that the encapsulator know how to properly apply ECMP to the MPLS
> packet that is being encapsulated.  That is, compute the hash that would be
> used to apply ECMP to the MPLS packet, and then map from that hash to a UDP
> source port.
>
> E.g., two MPLS packets with the same entropy label would get the same UDP
> source port, two MPLS packets with no entropy label but containing the same
> TCP flow would get the same source port, etc.
>
> Do you think there is a problem here?
> _______________________________________________
> mpls mailing list
> mpls@ietf.org
> https://www.ietf.org/mailman/listinfo/mpls