[Pals] draft-ietf-pals-ethernet-cw and enhanced heuristics

[Stewart Bryant <[e-mail]>]

The following has just been drawn to my attention:

https://www.juniper.net/documentation/en_US/junos/topics/concept/mpls-encapsulated-payload-load-balancing-overview.html

The writer contacted the authors, but not the list, and stated that this 
heuristic results in similar behaviour to that which we are trying to 
address in this draft, but that the draft as written does not address 
this problem.

I have asked the writer to post to the PALS list or let me do it for them.

In such implementations, even requiring the CW does not address the 
issue that is being reported to us.

Looking at RFC4385 it says:

   If a PW is sensitive to packet misordering and is being carried over
    an MPLS PSN that uses the contents of the MPLS payload to select the
    ECMP path, it MUST employ a mechanism that prevents packet
    misordering.  A suitable mechanism is the PWMCW described in Section
    3 for data, and the PWACH described in Section 5 for channel-
    associated traffic.

PWMCW is a reference to the use of the CW  starting with 0000.

I think that the implication in RFC4385 is that implementations should 
not proceed past the CW. That was certainly my intention when I wrote 
the text in 2005. I must say that whilst I knew of implementations that 
tried to perform ECMP on IP payloads, I was not aware of implementations 
that would attempt to glean that the PW was Ethernet and that its 
payload was IP and then perform five-tuple ECMP on the Ethernet payload.

I am wondering if we need to address this, and if so how to proceed.

One way would be to add text quoting RFC4385 and noting the use of the 
control word only prevents incorrect ECMP in implementations that cease 
further packet analysis on encountering a CW. Implementations that do 
packet analysis beyond the CW should expect unpredictable results, and 
therefore as noted in RFC4385 MUST employ an alternate mechanism to 
prevent packet misordering.

We could leave it there or we could suggest that using RFC6391 or 
RFC6790 with no further heuristics is a safe way to achieve ECMP.

- Stewart