Re: [tsvwg] David Black's WGLC review of draft-ietf-tsvwg-ecn-encap-guidelines-12

[Bob Briscoe <[e-mail]>]

David,

I've addressed all your points inline (and RS's and AMcG's points in 
recent emails).

I've edited my local copy of the draft as I describe. I'll submit it in 
a few hours time. I'm off to bed now, and in the morning I just want to 
double-check I haven't missed anyone's comments before posting it. That 
might also allow you, my co-author or anyone else enough time to come 
back on any of my responses below before I post...


On 09/05/2019 01:10, Black, David wrote:
>
> First of all, this is a well-written and comprehensive draft on ECN 
> encapsulation design guidelines.
>
> I found no major issues, and all of the minor issues are truly minor – 
> some border on editorial,
>
> and all can be dealt with via relatively minor edits..
>
> --- Minor ---
>
> [A] Need an explicit statement of exactly what the update to RFC 3819 is,
>
> preferably in a section whose name uses "RFC 3819" and appears in the
>
> table of contents.  Do this before some AD tells you to ;-).  This is
>
> also the only thing that idnits found that needs attention.
>

RFC 3819 contained a paragraph of brief musings on how subnetworks might 
handle ECN. So ecn-encap completely replaces that para with a whole 
document. I hope the new text quoted below is sufficient to explain that 
there is no specific section that updates RFC 3819. Rather the whole 
document does:

Copied sentence below from end of intro to abstract:
"This document updates the advice to subnetwork designers about ECN in 
RFC 3819."

Expanded sentence at the end of Intro to:
"This document updates and completely replaces the brief paragraph of 
advice to subnetwork designers about ECN in Section 13 of [RFC3819] (BCP 
89)."

> [B] 1. Introduction, first paragraph, last sentence
>
> The phrase "egress at any layer" is too general, hence incorrect (e.g.,
>
> layer 7 was not intended to be included) - "lower layer" needs to be
>
> used for clarity.
>
> OLD
>
>    If an egress at any layer is not ECN-aware, or if the
>
>    ultimate receiver or sender is not ECN-aware, congestion needs to be
>
>    indicated by dropping a packet, not marking it.
>
> NEW
>
>    If a lower layer header that may contain ECN congestion indications is
>
>    removed by an egress that is not ECN-aware, or if the ultimate 
> receiver or
>
>    sender is not ECN-aware, then lower layer congestion needs to be 
> indicated
>
>    by dropping a packet, not marking it.
>
> END
>
Thx. Used verbatim (except s/an egress/a subnet egress/).

> [C] 1.1 Scope
>
> Second paragraph and the two-bullet list imply that RFC 4774 states 
> that use
>
> of ECT(1) to signal alternative ECN semantics is a best current practice.
>
> That's not correct, as RFC 4774 covers only the first bullet on use of 
> DSCP,
>
> not the second bullet on use of ECT(1), as that second bullet is based 
> solely
>
> on RFC 8311. Some rewriting is in order to disconnect the second 
> bullet from
>
> RFC 4774.
>
Done:
"There are two main examples for how alternative ECN semantics have been 
defined in practice:
     o RFC 4774 suggests...
     o RFC 8311 suggests..."

> Based on this, all other places where RFC 4774 is mentioned or cited 
> should
>
> be checked to determine whether RFC 8311 also ought to be mentioned or 
> cited.
>
Done.

The one other occurrence of RFC4774 says:
     "The guidelines are also designed to ensure compliance with the 
more general best current practice for the design of alternate ECN 
schemes given in [RFC4774]..."

Added: "...and extended by [RFC8311]"


> [D] Page 5, first complete paragraph
>
>    Alternative semantics for the ECN field can be defined to depend on
>
>    the traffic class indicated by the DSCP.  Therefore correct
>
>    propagation of congestion signals could depend on correct propagation
>
>    of the DSCP between the layers.
>
> This also depends on correct propagation across diffserv boundaries - that
>
> and the potential damage wrought by bleaching DSCPs to zero at such 
> boundaries
>
> both deserve mention here.
>
Done; added:
"Similarly, if the DSCP is wiped at the boundary between Diffserv 
domains, the special ECN semantics would also be lost."


> [E] 1.1 Scope, end of section
>
>    In the feed-backward mode, propagation of
>
>    congestion signals for multicast and anycast packets is out-of-scope
>
>    (because it would be so complicated that it is hoped no-one would
>
>    attempt such an abomination).
>
> Much as I may agree with it :-), the parenthetical remark is inappropriate
>
> for an archival RFC, and hence needs to be removed.
>
When doing transport area reviews, I have criticised other authors for 
just ruling stuff out of scope without justification or explanation of 
the implications. So rather than delete, I've toned this down:

"... (because the complexity would make it unlikely to be attempted)."

BTW, for archival documents IMO light-heartedness is no less appropriate 
than unrelenting blandness.

> [F] 2. Terminology
>

>    ECN-PDU:  A PDU that is part of a feedback loop within which all the
>
>       nodes that need to propagate explicit congestion notifications
>
>       back to the Load Regulator are ECN-capable.
>
> That first sentence is too broad as it includes both the forward and
>
> reverse directions of the feedback loop, whereas only the forward
> direction is intended (FWIW, the definition of PDU does not suffice to
>
> narrow this to the forward direction only).  For example, a TCP packet
>
> with the ECE flag set in the TCP header and Not-ECT in the ECN field in
>
> the IP header would be an ECN-PDU under this definition, which is not
>
> what was intended.  The immediately following definition of Not-ECN-PDU
>
> has the same problem.
>
The point of building on the OSI term 'PDU' was because a PDU is defined 
by the layer that is stated as relevant. By the OSI definition, a PDU at 
a certain layer does not include lower layer headers that encapsulate it.

So, in the example you give of a TCP packet with ECE set, that is a TCP 
PDU. By OSI definition, a TCP PDU does not include the IP header that 
encapsulates it. The IP header is only relevant when considering it as 
an IP PDU. And then the TCP PDU within it is just payload (the SDU that 
was passed to IP) and not part of the protocol of the IP PDU.

> The narrowing concept that needs to be added is that if the PDU carries
>
> a congestion indication, then that congestion indication participates in
>
> such a feedback loop,
>
Yes, valid point. Thank you.

    ECN-PDU:  A PDU that is part of a feedback loop within which all the

       nodes that need to propagate *the* explicit congestion notification

*signal in the PDU* back to the Load Regulator are ECN-capable *of doing 
so*.


> and even that is subtle when there are multiple
>
> possible congestion indication locations.  An example of the subtlety is
>
> that the same PDU may be an ECN-PDU in feed-up-and-forward mode because
>
> the layer 3 and 4 protocols support ECN but may be a Not-ECN-PDU for
>
> feed-forward-and-up mode because the L2 decapsulator ignores and discards
>
> L2 congestion indications.
>
I think the bold words in the phrase "nodes that *need to* propagate" 
deals with that subtlety.

In your example, if the subnet is propagating the PDU's ECN signal in 
feed-up-and-forward mode and the L2 decapsulator doesn't propagate ECN, 
it's not relevant because it also doesn't *need to* propagate ECN. 
Whereas, if the subnet propagates ECN in feed-forward-and-up mode, the 
decap does *need to*.

I've also extended the above correction into the next definition:

    Not-ECN-PDU:  A PDU that is part of a feedback-loop within whichsome  *at least one*
       node necessary to propagate*the*  explicit congestion notification*signal in the PDU *  
       back to the load regulatorare  *is*  notECN-capable*of doing so*.


> [G] 2. Terminology
>
>    Congestion Baseline:  The location of the function on the path that
>
> initialised the values of all congestion notification fields in a
>
>       sequence of packets, before any are set to the congestion
>
> experienced (CE) codepoint if they experience congestion further
>
> downstream.  Typically the original data source at layer-4.
>
> That's counter-intuitive - a baseline is a level, not a location. I've
>
> tagged this as a minor issue as opposed to editorial because it causes
>
> severe confusion in the only place that this term is used, namely item 3
>
> in Section 4.3. That item 3 is about a Monitoring Baseline which is
>
> a level not a location.
>
> The Congestion Baseline term should be removed, and item 3 in Section
>
> 4.3 revised accordingly.  The term "Congestion Baseline" is used twice
>
> there, but only the latter of those two uses needs to be retained and
>
> expanded after deletion of this definition.
>
You're right. I've done as you suggest.

I've also nearly completely re-written the second half of the section, 
because I had temporarily forgotten the rationale, and even I couldn't 
understand what the rationale was after reading my own words.

> [H] 3. Modes of Operation
>
> Feed-Up-and-Forward:  A lower layer switch feeds-up congestion
>
> notification directly into the ECN field in the higher layer
>
>          (e.g. IP) header, irrespective of whether the node is at the
>
>          egress of a subnet.
>
> Remove "the ECN field in" as that's too restrictive, even though that's
>
> a common case, and move it into the example – “(e.g., ECN field in the
>
> IP header)" where "header" moved inside the parentheses.
>
Done.


> [I] 3.3. Feed-Backward Mode
>
> Please make it clear that the critique of the ATM ABR relative rate
>
> control mechanism also applies to Ethernet QCN, as that's a more
>
> current example.  This should be connected to the QCN discussion at
>
> the end of Section 4.2.
>
I didn't want to imply any criticism of QCN, which always made it clear 
it was only applicable to a single subnet.

However, I can now see the need to have this written down, given certain 
companies are giving incentives to churn out Frankenpatents (any random 
combination of existing ideas, no matter whether useful or correct). So 
I've kept ATM as the example, but added at the end:

    QCN [IEEE802.1Qau] would suffer from similar problems if extended to
    multiple subnets.  However, from the start QCN was clearly stated as
    solely applicable to a single subnet (see Section 6).


> [J] 4.3 Encapsulation Guidelines
>
> Resolving minor issue [G] requires some edits to item 3 in this section:
>
> - Remove "(the Congestion Baseline)" from the third paragraph.
>
> - Rewrite start of fourth paragraph:
>
> OLD
>
>        Most information can be extracted if the Congestion Baseline is
>
> standardised at the node that is regulating the load (the Load
>
> Regulator--typically the data source).
>
> NEW
>
>        More information can be extracted if the protocol specification
>
>        requires that congestion fields be initialized to indicate no
>
> congestion only by the node that is regulating the load (the
>
>        Load Regulator--typically the data source).
>
> END
>
See minor issue [G] above - already completely rewritten this text.


> [K] 12.2 Informative References
>
> Please check the references to 802.1Qah and 802.1Qau with IEEE.  The
>
> correct references to use now may be the latest version of 802.1Q with
>
> specific functionality and/or clauses identified.
>
Done.
802.1Qau -> 802.1Q Sections 30--33
802.1ah -> 802.1Q (as a whole - too scattered to call out particular 
sections).

Pat Thaler having retired, I no longer have a co-author who can advise 
on IEEE citation correctness.
But the above is good enough.

[I knew someone would point that out - I should have dealt with it but, 
even tho I purportedly have free access to the full rolled-up text of 
the current standard (IEEE802.1Q-2018), I've always had trouble viewing 
it. The IEEE pay-wall insists on displaying the free copy in my browser 
in a way that it can't be downloaded, but in Firefox the activity bar 
spins but nothing ever appears. After much gnashing of teeth, I 
discovered it works OK in Chrome.]

> --- Editorial/Nits ---
>
> Abstract, last sentence
>
>    Following these guidelines should assure
>
>    interworking between new lower layer congestion notification
>
>    mechanisms, whether specified by the IETF or other standards bodies.
>
> between new -> among IP layer and
>
Yup.
>
> 1. Introduction, first paragraph
>
> Suggest removing use of the word "buffer" - "lower layer" suffices in
>
> all of the places where this word is used in this paragraph.
>
Not as simple as that. I've guessed that the idea of a buffer doing 
marking jars with you, so I've addressed the two occurrences of 'buffer' 
as follows:

1.   When a lower layer buffer drops a packet

This was intended to introduce the context as congestion loss, not 
transmission loss. So left unchanged.

2. In
    contrast, when a lower layer marks a packet with ECN, the marking
    needs to be explicitly propagated up the layers.  The same is true if
    a buffer marks the outer header of a packet that encapsulates inner
    tunnelled headers.

The second sentence is about tunnels, not lower layers.
So, I've introduced AQM into both sentences:

    In
    contrast, when active queue management (AQM) at a lower layer marks a
    packet with ECN, the marking needs to be explicitly propagated up the
    layers.  The same is true if AQM marks the outer header of a packet
    that encapsulates inner tunnelled headers.


> 1. Introduction, second paragraph
>
>    The purpose of this document is to guide the addition of congestion
>
>    notification to any subnet technology or tunnelling protocol, so that
>
>    lower layer equipment can signal congestion explicitly and it will
>
>    propagate consistently into encapsulated (higher layer) headers,
>
>    otherwise the signals will not reach their ultimate destination.
>
> Suggest: "equipment" -> "functionality"
>
Sorry, the word 'functionality' gets to me like finger-nails on a 
blackboard.
     https://www.dailywritingtips.com/is-that-even-a-word/

I've used s/equipment/AQM algorithms/


> Page 5, first full paragraph:
>
>    If a particular protocol design chooses to contradict a
>
>    'SHOULD (NOT)' given in the advice below, it MUST include a sound
>
> justification.
>
> chooses to contradict -> chooses not to follow
>
Yup
>
> 1.1. Scope, first paragraph
>
>    This document only concerns wire protocol processing of explicit
>
>    notification of congestion.
>
> Remove "wire"
>
I think that loses the intended sense in too much ambiguity.
'Protocol processing' could include the AQM algorithms, which this 
sentence is trying to exclude.

> Section 3.1, second paragraph
>
>    In these cases, ECN may best be supported by standardising explicit
>
>    notification of congestion into the lower layer protocol that carries
>
>    the data forwards.  It will then also be necessary to define how the
>
>    egress of the lower layer subnet propagates this explicit signal into
>
>    the forwarded upper layer (IP) header.  It can then continue forwards
>
>    until it finally reaches the destination transport (at L4).  Then
>
>    typically the destination will feed this congestion notification back
>
>    to the source transport using an end-to-end protocol (e.g.  TCP).
>
>    This is the arrangement that has already been used to add ECN to IP-
>
>    in-IP tunnels [RFC6040], IP-in-MPLS and MPLS-in-MPLS [RFC5129].
>
> The referents of "It" at the start of the second and third sentences are
>
> unclear and different.  Suggested rephrasings:
>
> It will then also be necessary to define ->
>
>      This necessitates also specifying
>
> It can then continue -> This signal continues
>
Good.
>
> Section 3.1, first paragraph after Figure 1
>
> Cite one or more of the 3GPP references for GTP tunnels, as this is the
>
> first occurrence of that concept in this draft.
>
Yup
>
> Section 3.1, second paragraph after Figure 1
>
> Cite a Frame Relay reference for the FECN bit..  Moving the [Buck00]
>
> citation up to immediately follow "Frame Relay" in the first line
> suffices.
>
Good.
>
> Section 3.2 first paragraph
>
>    These are Ethernet switches that bury into the Ethernet payload ...
>
> bury -> burrow
>
> [Spelling checker is missing “read user’s mind” feature :-)]
>
I've used dig.

Hard to read user's mind. Easier to tell user what to think ;)

> Section 3.4
>
> The second paragraph could be improved by citing references for fat-tree
>
> and Clos network topologies.
>
Done.

> Section 4.1, second paragraph
>
>    Therefore section 4 of
>
> [I-D.ietf-tsvwg-rfc6040update-shim] requires network operators to
>
>    configure the ingress of a non-ECN tunnel so that it zeros the ECN
>
>    field in the outer IP header.
>
> non-ECN tunnel -> tunnel that does not support ECN
>
OK
>
> Section 4.1, third paragraph
>
>    Even
>
>    if the shim(s) encapsulate a L2 header, it is often possible to find
>
>    an inner IP header within the L2 header
>
> Latter instance of "L2 header" should be "L2 PDU"
>
Good.
>
> Section 4.1, last paragraph
>
>    Instead a companion specification
>
> [I-D.ietf-tsvwg-rfc6040update-shim] has been prepared that has
>
>    sufficient standards track status to update standards track
>
>    protocols.
>
> sufficient -> the appropriate
>
Done.
>
> Authors Addresses
>
> Please check these - Pat Thaler's contact info probably needs to be 
> updated.
>
In progress.



Bob
>
> Thanks, --David
>
> ----------------------------------------------------------------
>
> David L. Black, Senior Distinguished Engineer
>
> Dell EMC, 176 South St., Hopkinton, MA 01748
>
> [phone] *New *   Mobile: [phone]
>
> [e-mail] <mailto:[e-mail]>
>
> ----------------------------------------------------------------
>

-- 
________________________________________________________________
Bob Briscoehttp://bobbriscoe.net/