Re: [bess] Signaling Control Word in EVPN

["Andrew G. Malis" <[e-mail]>]

James,

It's much harder to mandate use of EL than the CW for several reasons:
- CW implementation is much more common than EL implementation
- PWs and/or EVPN are rarely the only traffic in an MPLS traffic tunnel,
rather, they will be multiplexed with other MPLS-based applications that
are using the traffic tunnel to reach a common destination. Thus, by using
the CW, you can disable ECMP only for those MPLS packets that cannot
tolerate reordering.

That said, I'm also concerned that because of the existing text in 7432,
implementations may not be using the CW even for P2P EVPN.

And we still don't have a good answer for Muthu's original question. :-)

Cheers,
Andy


On Tue, Oct 9, 2018 at 8:49 AM James Bensley <[e-mail]> wrote:

> On Tue, 9 Oct 2018 at 11:29, Alexander Vainshtein
> <[e-mail]> wrote:
> > I fully agree that reordering due to lack of the CW happens, and that
> usage of the CW is the right way to eliminate that.
>
> Hi Sasha,
>
> If I may rudely interject here. You have to be careful with your
> wording. The PWMCW doesn't eliminate reordering.
>
> https://tools.ietf.org/html/draft-ietf-pals-ethernet-cw-07#section-4:
>
> 4.  Recommendation
>
>    The ambiguity between an MPLS payload that is an Ethernet PW and one
>    that is an IP packet is resolved when the Ethernet PW control word is
>    used.  This document updates [RFC4448] to state that where both the
>    ingress PE and the egress PE support the Ethernet pseudowire control
>    word, then the CW MUST be used.
>
> The CW doesn't fix the problem in every case, but it does fix it in most
> cases.
>
> The only way to truly fix this problem is to use entropy labels. The
> problem with the CW draft is that it only recommends entropy labels or
> FAT labels when ECMP is required, not when it isn't require *but*
> exists anyway within the packet switched core:
>
>    Where the application of ECMP to an Ethernet PW traffic is required,
>    and where both the ingress and the egress PEs support [RFC6790]
>    (Entropy Label Indicator/Entropy Label (ELI/EL)) or both the ingress
>    and the egress PEs support [RFC6391] (FAT PW), then either method may
>    be used.
>
> For this reason my personal opinion is that only EL should be
> recommended by the EVPN draft. I don't think a solution that doesn't
> completely fix the problem (PWMCW) should be recommend when we have
> one that does fully fix the problem (El/ELI or ELC/RLD in SR).
>
> > Or should we agree that consistent usage of the CW in the EVPN
> encapsulation has to be delegated to some network-wide management
> mechanism/LSO? For the reference, the latest (expired) version of the EVPN
> YANG data model does not include any attributes that indicate usage or
> non-usage of the CW in the encapsulation...
>
> This is perhaps the more difficult question ^ what to recommend in the
> case that EL isn't supported? You could recommend that all devices
> should be configured to *not* hash on the VPN payload and only the
> label stack, then there can be no mistaking what the VPN contents are
> (Ethernet/IPv4/IPv6/other) but, some devices might not support that.
> Alternatively, you could recommend that the CW is used only when EL
> isn't available but, some devices might not support that. I'll leave
> this up to you guys as you're the experts here, I just wanted to point
> out the wording issue above regarding CW and EL and "fixing"
> reordering.
>
> Cheers,
> James.
>