Re: [Idr] AD Review of draft-ietf-idr-bgp-optimal-route-reflection-22

[Alvaro Retana <[e-mail]>]

Ping!

On March 18, 2021 at 3:32:01 PM, Alvaro Retana ([e-mail])
wrote:


Dear authors:


Thank you for your work on this document.  I know that the draft and the
related ideas have been around for a long time.

I have a couple of significant issues that I want to highlight upfront, and
many comments inline (see below).  In general, my focus is to make sure
that the specification of this mechanism is clear, not just to the
experienced reader, but to the casual one as well.  I believe that all the
issues should be easy to resolve.

(1) Terminology.  Please use the terminology already defined in
rfc4271/rfc4456 instead of creating new one by indirection.  My first
inline comment is about the use of "best path", which is not what rfc4271
uses -- there are other occurrences later on.

(2) Deployment Considerations.  §6 says that a compliant implementation is
one that allows the operator to configure "a logical location from which
the best path will be computed, on the basis of either a peer, a peer
group, or an entire routing instance".  Please spend some time providing
guidance so an operator can decide what best works for their network.  [I
pointed below to other places there operational guidance would be ideal.]

(3) I don't think that §5 (CPU and Memory Scalability) and Appendix A
(alternative solutions with limited applicability) should be part of this
document.  To start, comparisons are never good and the work has already
been through the WG so there's no need to keep justifying it by comparing
it to other work.  Both sections contain several claims (e.g. "extra cost
in terms of CPU", "implemented efficiently", "expected to be higher but
comparable", "large amount of BGP state and churn", "result in tens of
paths for each prefix") that don't have references, and could be considered
subjective because they are clearly relative and can change depending on
the specific deployment and configuration of the network.  I am not saying
that the claims are false, simply requesting to take these two sections
out.


Thanks!

Alvaro.


[Line numbers from idnits.]

...
17 Abstract

19   This document defines an extension to BGP route reflectors.  On route
20   reflectors, BGP route selection is modified in order to choose the
21   best path from the standpoint of their clients, rather than from the
22   standpoint of the route reflectors.  Multiple types of granularity
23   are proposed, from a per client BGP route selection or to a per peer
24   group, depending on the scaling and precision requirements on route
25   selection.  This solution is particularly applicable in deployments
26   using centralized route reflectors, where choosing the best route
27   based on the route reflector IGP location is suboptimal.  This
28   facilitates, for example, best exit point policy (hot potato
29   routing).

[major] "best path"  rfc4271 doesn't use the term "best path".  The
terminology used in this document should, at the very least, match what the
base spec defines.  Let's please not create new terminology by indirection
using the "definition of terms" section.


[minor] "proposed"  This document is in line to be come an RFC; we're far
beyond proposing things.  There are a couple of places later on where this
same wording is used.  Please change it to specified or at least described
(the latter probably fits best in this specific case).


[nit] s/from a per client BGP route selection or to a per peer group/from
per client BGP route selection or per peer group


[nit] s/precision requirements on route selection./precision requirements.


[nit] s/route reflector IGP location/route reflector's IGP location


[major] "IGP location"  Before I forget -- please clearly define (not in
the Abstract of course) what an "IGP location" is.


...
70   1.  Definitions of Terms Used in This Memo  . . . . . . . . . . .   2
71   2.  Introduction  . . . . . . . . . . . . . . . . . . . . . . . .   3

[nit] It would be nice if the Introduction came up before the terminology.


...
91 1.  Definitions of Terms Used in This Memo

[minor] To avoid repeating some terms, it is good practice to indicate
which other RFCs the reader should be familiar with.  In this case rfc4271
and rfc4456 come to mind.


...
132 2.  Introduction
...
153   Section 11 of [RFC4456] describes a deployment approach and a set of
154   constraints which, if satisfied, would result in the deployment of
155   route reflection yielding the same results as the IBGP full mesh
156   approach.  This deployment approach makes route reflection compatible
157   with the application of hot potato routing policy.  In accordance
158   with these design rules, route reflectors have traditionally often
159   been deployed in the forwarding path and carefully placed on the POP
160   to core boundaries.

[nit] s/have traditionally often/have often
Redundant...


162   The evolving model of intra-domain network design has enabled
163   deployments of route reflectors outside of the forwarding path.
164   Initially this model was only employed for new address families, e.g.
165   L3VPNs and L2VPNs, however it has been gradually extended to other
166   BGP address families including IPv4 and IPv6 Internet using either
167   native routing or 6PE.  In such environments, hot potato routing
168   policy remains desirable.

[minor] "new address families, e.g. L3VPNs and L2VPNs"  These are not the
name of the AFs.  Maybe call them new services.


[minor] "IPv4 and IPv6 Internet"  The name of the AF is IP/IP6; again, you
probably mean service/application.


[nit] "native routing or 6PE"  Not 4PE applications?  ;-)   It seems to me
that you can save some ink by ending the sentence after "Internet".


...
194 3.  Modifications to BGP Best Path selection

196   The core of this solution is the ability for an operator to specify
197   the IGP location for which the route reflector should calculate
198   routes.  This can be done on a per route reflector basis, per peer/
199   update group basis, or per peer basis.  This ability enables the
200   route reflector to send to a given set of clients routes with
201   shortest distance to the next hops from the position of the selected
202   IGP location.  This provides for freedom of route reflector physical
203   location, and allows transient or permanent migration of this network
204   control plane function to an arbitrary location.

[major] "ability for an operator to specify the IGP location for which the
route reflector should calculate routes.  This can be done on a per route
reflector basis, per peer/update group basis, or per peer basis."

When should an operator choose per RR vs per peer group vs per peer?
Choice is good, but providing guidance for the deployment is much better.
Please consider adding text to §6 about the considerations to chose one or
the other.


[minor] "peer/update group"  Where is this defined?  Is there a reference
to a non-implementation-specific definition?  Can it be called "a group of
peers"?  Maybe you also want to include it in the definitions in §1.


...
215   o  it can, and usually will, have a different position in the IGP
216      topology, and

[] "usually will"   The position in the topology will *always* be
different!


...
222   This document defines, on BGP Route Reflectors [RFC4456], two changes
223   to the BGP Best Path selection algorithm:

[nit] s/defines, on BGP Route Reflectors/defines, for BGP Route Reflectors


...
235   A route reflector can implement either or both of the modifications
236   in order to allow it to choose the best path for its clients that the
237   clients themselves would have chosen given the same set of candidate
238   paths.

[major] Please provide guidance for the operator to consider then one or
both should be used in their network.


240   A significant advantage of these approaches is that the route
241   reflector clients do not need to run new software or hardware.

[nit] s/do not need to run new software or hardware./do not need to be
modified.


243 3.1.  Best Path Selection from a different IGP location

245   In this approach, optimal refers to the decision made during best
246   path selection at the IGP metric to BGP next hop comparison step.  It
247   does not apply to path selection preference based on other policy
248   steps and provisions.

[major] Please be specific about what is the "IGP metric to BGP next hop
comparison step".  There is no step with that name, or even a string match
in rfc4271.  Later you talk about the step e) tie-breaker -- don't wait to
be specific!


...
263      e) Remove from consideration any routes with less-preferred
264      interior cost.  The interior cost of a route is determined by
265      calculating the metric from the selected IGP location to the
266      NEXT_HOP for the route using the shortest IGP path tree rooted at
267      the selected IGP location.

[major] Note that the specification talks about "interior cost", but the
descriptions in this document uses "IGP cost" and "IGP metric" to refer to
the same thing.  Please be consistent with existing terminology!


[major] Related to the definition of "IGP location" and its configuration.
How does the IGP location (and thus the calculation of the interior cost,
as above) change when the configuration is done per peer/update groups
instead than per peer?  This is another point where operators could benefit
from guidance (for the peer/update-group and whole RR cases, of course).


...
275   The configuration of the IGP location is outside of the scope of this
276   document.  The operator may configure it manually, an implementation
277   may automate it based on heuristics, or it can be computed centrally
278   and configured by an external system.

[nit] s/outside of the scope/outside the scope

[major] There are several places later on that talk about the
configuration, even making it a requirement for compliance (§6).  I think
that you don't mean the configuration itself, but the way in which the
configuration is instantiated in the router (which could be considered an
implementation detail).  As written, the text is not clear.


280   This solution does not require any change (BGP or IGP) on the
281   clients, as all required changes are limited to the route reflector.

[] This was also claimed before.


283   This solution applies to NLRIs of all address families that can be
284   route reflected.

[] Are there AFs that cannot be reflected?  Just wondering if you really
need to say this.


286 3.1.1.  Restriction when BGP next hop is BGP prefix

288   In situations where the BGP next hop is a BGP prefix itself, the IGP
289   metric of a route used for its resolution SHOULD be the final IGP
290   cost to reach such next hop.  Implementations which can not inform
291   BGP of the final IGP metric to a recursive next hop SHOULD treat such
292   paths as least preferred during next hop metric comparison.  However
293   such paths SHOULD still be considered valid for best path selection.

[major] The recursive resolution is already covered in rfc4271
(§5.1.3/§9.1.2.1).  Why isn't that enough?

There seems to be a difference: if an implementation "can not inform BGP of
the final IGP metric to a recursive next hop...SHOULD still be considered
valid for best path selection."  rfc4271 treats these routes as
unresolvable.  Do you want to consider unresolvable routes?

Finally, there are a lot of SHOULDs in this paragraph.  Under which
conditions is it ok to not perform the actions?  IOW, why are these actions
recommended and not required?


295 3.2.  Multiple Best Path Selections

297   BGP Route Reflector as per [RFC4456] runs a single best path
298   selection.  Optimal route reflection may require calculation of
299   multiple best path selections or subsets of best path selection in
300   order to consider different IGP locations or BGP policies for
301   different sets of clients.

[] It hasn't been mentioned before, but the talk about policy made me think
about this:   Can a client be present on different sets?  How is that
handled in terms of BGP sessions?  Do we need multiple sessions?  Or is
add-path required?


303   If the required routing optimization is limited to the IGP cost to
304   the BGP Next-Hop, only step e) as defined [RFC4271] section 9.1.2.2,
305   needs to be duplicated.

[major] I'm not sure if this paragraph is referencing §3.1 (where step e)
was just modified), or or you're saying that for each subset (from the last
paragraph) only step e) is considered, or something else.  ??


307   If the routing optimization requires the use of different BGP
308   policies for different sets of clients, a larger part of the decision
309   process needs to be duplicated, up to the whole decision process as
310   defined in section 9.1 of [RFC4271].  This is for example the case
311   when there is a need to use different policies to compute different
312   degree of preference during Phase 1.  This is needed for use cases
313   involving traffic engineering or dedicating certain exit points for
314   certain clients.  In the latter case, the user MAY specify and apply
315   a general policy on the route reflector for a set of clients.  For a
316   given set of clients, the policy SHOULD in that case allow the
317   operator to select different candidate exit points for different
318   address families.  Regular path selection, including IGP perspective
319   for a set of clients as per Section 3.1, is then applied to the
320   candidate paths to select the final paths to advertise to the
321   clients.

[] Similar comment as before...   The use of "duplicated" is confusing me.


[major] "the user MAY specify..."  s/MAY/may   It seems to me that you're
simply stating a fact (the user can do this), and not specifying an
optional behavior.


[major] "...the policy SHOULD in that case allow the operator to select
different candidate exit points for different address families."  There's
no interoperable action that "the policy" can execute.  It sounds as if
you're recommending that specific knobs ("allow the operator to...") be
implemented.  Please reword.


[] You use "IGP perspective" I guess as equivalent to "IGP location" -- is
that right?


323 4.  Implementation considerations

325 4.1.  Likely Deployments and need for backup

[nit] Do we really need a sub-section if all the text for §4 is in it?


327   With IGP based optimal route reflection, even though the IGP location
328   could be specified on a per route reflector basis or per peer/update
329   group basis or per peer basis, in reality, it's most likely to be
330   specified per peer/update group basis.  All clients with the same or
331   similar IGP location can be grouped into the same peer/update group.
332   An IGP location is then specified for the peer/update group.  The
333   location is usually specified as the location of one of the clients
334   from the peer group or an ABR to the area where clients are located.
335   Also, one or more backup locations SHOULD be allowed to be specified
336   for redundancy.  Implementations may wish to take advantage of peer
337   group mechanisms in order to provide for better scalability of
338   optimal route reflector client groups with similar properties.

[major] "IGP based optimal route reflection"

This is the first time you use "IGP based optimal route reflection"; I
peeked forward and see that §5 also mentions "policy based optimal route
reflection".  But you didn't define them anywhere -- in fact, the
description before now focuses on the IGP location, which makes me think
about "IGP based".

I could guess that the "policy based" version is related to §3.2, but (1)
no one should have to guess (!), and (2) the end of that section says that
"IGP perspective...is then applied", making it also "IGP based".  ??

Please either define these types of ORR earlier in the text, or (better
yet) don't use the terms.  Instead, and given that the names are used just
a couple of times, change "policy based" to "the case where a policy is
applied"...


[major] "one or more backup locations SHOULD be allowed to be specified for
redundancy"

(1) §3.1 says that the "configuration of the IGP location is outside of the
scope of this document".  If that is true, then you can't recommend
anything related to the configuration.

(2) If there were "backup locations", how would they be used?


[minor] This is the only time that "optimal route reflector client groups"
is used.  I had been assuming that a group of clients would correspond to a
specific peer/update group (as mentioned many times before).  Is there a
difference between a "client group" and grouping clients to correspond to a
peer/update group?


[] "for better scalability of optimal route reflector client groups"  What
exactly does this mean?  I asked before about references/definitions for
peer/update groups; this point may be related.


...
373 6.  Advantages and Deployment Considerations

375   The solutions described provide a model for integrating the client
376   perspective into the best path computation for route reflectors.
377   More specifically, the choice of BGP path factors in either the IGP
378   cost between the client and the nexthop (rather than the IGP cost
379   from the route reflector to the nexthop) or other user configured
380   policies.

[] "solutions"?   I only see one, where the "second" one simply adds
policy, which is a pervasive tool in BGP.


382   The achievement of optimal routing relies upon all route reflectors
383   learning all paths that are eligible for consideration.  In order to
384   satisfy this requirement, path diversity enhancing mechanisms such as
385   BGP add-path [RFC7911] may need to be deployed between route
386   reflectors.

[mayor] "path diversity enhancing mechanisms...may need to be deployed
between route reflectors"  Given that this is the Deployment Considerations
section, please provide guidance on when these mechanisms are needed, and
when they're not.


[minor] "path diversity enhancing mechanisms such as BGP add-path
[RFC7911]"  Just out of curiosity, what other mechanisms are you thinking
of?  Are there deployment considerations, when using ORR, to prefer one?


388   Implementations considered compliant with this document allow the
389   configuration of a logical location from which the best path will be
390   computed, on the basis of either a peer, a peer group, or an entire
391   routing instance.

[] I guess that "logical location" is another name for IGP location...


[major] §3.1 says that the "configuration of the IGP location is outside of
the scope of this document".  If that is true, then the configuration
cannot be a consideration for compliance.

A better wording for compliance would be: "An implementation MUST allow the
configuration..."


393   These solutions can be deployed in traditional hop-by-hop forwarding
394   networks as well as in end-to-end tunneled environments.  In networks
395   where there are multiple route reflectors and hop-by-hop forwarding
396   without encapsulation, such optimizations SHOULD be enabled in a
397   consistent way on all route reflectors.  Otherwise, clients may
398   receive an inconsistent view of the network, in turn leading to
399   intra-domain forwarding loops.

[major] "optimizations SHOULD be enabled in a consistent way on all route
reflectors"

Is "a consistent way" different than "on all"?   s/.../optimizations SHOULD
be enabled on all route reflectors

When is it ok for all RRs not to be enabled with the optimizations?  IOW,
why is it recommended and not required?  Avoiding "intra-domain forwarding
loops" sounds like a good reason to require it.


...
406   As per above, these approaches reduce the amount of state which needs
407   to be pushed to the edge of the network in order to perform hot
408   potato routing.  The memory and CPU resources required at the edge of
409   the network to provide hot potato routing using these approaches is
410   lower than what would be required to achieve the same level of
411   optimality by pushing and retaining all available paths (potentially
412   10s) per each prefix at the edge.

[] This paragraph sounds very speculative and doesn't seem necessary.  In
line with §5.


414   The solutions above allow for a fast and safe transition to a BGP
415   control plane using centralized route reflection, without
416   compromising an operator's closest exit operational principle.  This
417   enables edge-to-edge LSP/IP encapsulation for traffic to IPv4 and
418   IPv6 prefixes.

[] "allow for a fast and safe transition"  Besides this statement sounding
like a marketing brochure, if you're going to talk about transition, please
talk about it in detail.  Please take a look at §2/rfc5706 and include
details about the transition and how it is "fast and safe".


420   Regarding Best Path Selection from a different IGP location, it
421   should be self evident that this solution does not interfere with
422   policies enforced above IGP tie-breaking in the BGP best path
423   algorithm.

[minor] Don't assume anything is self evident to anyone.

Suggestion (maybe more appropriate in §3.1)>
   The modification specified in Section 3.1 does not impact any previous
   considerations in the BGP Decision Process.


425 7.  Security Considerations

427   Similarly to [RFC4456], this extension to BGP does not change the
428   underlying security issues inherent in the existing IBGP [RFC4456].

[minor] No need to reference rfc4456 twice in the same sentence.


430   It however enables the deployment of base BGP Route Reflection as
431   described in [RFC4456] to be possible using virtual compute
432   environments without any negative consequence on the BGP routing path
433   optimality.

[] I'm not sure how this relates to the specification itself (platform
independent), but I'm sure the Security ADs will have a lot of questions
about the security of "virtual compute environments" -- and where the
details are included in this document.   IOW, that sounds like an
unnecessary assertion.


[major] "without any negative consequence"  The new functionality specified
in this document is having the RR run the client's selection for them (even
specific policy) -- there are risks related to the duplication of the
policy, the location of the RRs and the blind trust that the clients need
to have.  Or should the clients rerun the policy locally?

This behavior is not necessarily different from normal RR, but the
instantiation is different.   It would be very easy for a rogue RR to
propagate incorrect routing information to its clients -- specially if
policy is offloaded to them.

In the worst case the result would be a sub-optimal route (maybe even worse
than what the clients get today), but probably not a "real" security
issue.  It is probably a good idea to explain why it is not a security
risk.


435   This document does not introduce requirements for any new protection
436   measures, but it also does not relax best operational practices for
437   keeping the IGP network stable or to pace rate of policy based IGP
438   cost to next hops such that it does not have any substantial effect
439   on BGP path changes and their propagation to route reflection
440   clients.

[major] "best operational practices"  Like what?  It would be good to
provide (at least) informational references.


[major] "to pace rate of policy based IGP cost to next hops such that it
does not have any substantial effect on BGP path changes"   You will have
to explain this further because the operation of an IGP is not mentioned
anywhere else.

IMHO, there's no need to go into IGP operational details in this document.
There is nothing special/different about the operation of an IGP.


[End of Review]