Liyan,
On 07/07/2026 10:44, Liyan Gong wrote:
Dear All,
Thank you very much for the constructive feedback.
We apologize for the delayed response. We took some time to carefully
rethink our proposal and align it with the existing architecture.
After revisiting the Flex-Algo architecture (RFC 9350) and related
work (RFC 9843), we realized that *the core question is not only about
whether normalization is useful,**but also about where the
normalization logic should be defined and applied*. We tried to reason
from the layered design of Flex-Algo:
-*Link Layer:* Links advertise only their intrinsic, raw properties
(IGP metric, delay, loss, bandwidth). They should not be aware of, or
responsible for, any per-FA normalization logic.
- *FAD Layer: *The FAD defines the calculation rules for a given FA —
which Metric-Type to use, what constraints to apply, and how paths
should be computed.
We view metric normalization parameters (metric-step and
metric-offset) as *calculation policy at the FA**level*, not as link
configuration. Different FAs may have different optimization goals and
thus different normalization needs.
FAD was defined to synchronize the path calculation parameters.
Metric value is a local parameter, not subject to network wide
synchronization. Any metric adjustment, normalization or any other
manipulation is a local matter that can be done differently on every node,
So I disagree with your point of view.
thanks,
Peter
We have also revisited RFC 9843, which extends Flex-Algo to support
bandwidth-based metric calculation. We see this as a helpful
precedent: parameters like Reference Bandwidth are carried in the FAD
to define how metrics are calculated. Our proposal follows the same
pattern, extending this capability to other metric types such as delay
and loss.
Based on this reasoning, we have submitted an updated version (-01) of
the draft. We fully understand that the WG may still have reservations
about this approach, and we are genuinely open to further guidance. If
the WG feels it would be helpful, we would respectfully request an
opportunity to present this updated version at the upcoming IETF meeting.
Thank you again for your time and feedback and We look forward to your
thoughts.
Best Regards,
Liyan(on behalf of the co-authors)
----邮件原文----
*发件人：*Acee Lindem <[email protected]>
*收件人：*Peter Psenak <[email protected]>
*抄 送: *Liyan Gong <[email protected]>,Gyan Mishra
<[email protected]>,lsr <[email protected]>
*发送时间：*2025-11-08 23:27:59
*主题：*Re: [Lsr] draft-gl-lsr-metric-normalize/
Speaking as WG member:
I agree with Peter and have a couple additional concerns.
1. If metric normalization is to be performed in the IGP domain,
the LSA/LSP originator should
do it rather than having everyone else in the domain do it
for the originator's metrics.
2. The stated goal is to compute more ECMP paths. Did you
consider that some vendors
have implemented UCMP (Unequal Cost MultiPath)?
Thanks,
Acee
> On Nov 6, 2025, at 8:16AM, Peter Psenak wrote:
>
> Liyan,
> On 06/11/2025 10:40, Liyan Gong wrote:
>> Hi All,
>> hank you for your valuable comments.
>> To ensure we are fully aligned, we would like to further
clarify the considerations behind our proposal.
>> Please feel free to correct us if we have misunderstood any
aspect of your suggestion.
>> In designing our approach, we focused on addressing two key
aspects:
>> 1. **Parameter Consistency**: Parameters such as metric-step
and metric-offset need to be consistent across the network. We
observed that any inconsistency in static configuration could lead
to incorrect normalized metrics, which might in turn affect path
selection.
> there is no need for consistency of these parameters across the
network.
> If the configuration consistency is the problem you trying to
address, there are existing tools for that.
>> 2. **Parameter Optimization**: In dynamic network
environments, these parameters may need to be tuned. We think that
relying solely on manual reconfiguration in such cases could not
only be operationally costly and prone to errors, but also slow
down routing convergence, potentially extending periods of network
instability.
> similarly, FAD is not a configuration tool, there are existing
tools for that.
> thanks,
> Peter
>
>> These considerations motivated our proposal, which draws
inspiration from the established concept of automatic metric
calculation — an approach that addresses similar needs. We believe
a protocol-based extension could be a suitable way to resolve
these issues while staying consistent with existing mechanisms.
>> We welcome further discussion.
>>
>> Best Regards,
>> Liyan
>>
>> -----邮件原件-----
>> 发件人: Acee Lindem
>> 发送时间: 2025年11月6日 3:06
>> 收件人: Gyan Mishra
>> 抄送: Peter Psenak ; [email protected]
>> 主题: [Lsr] Re: draft-gl-lsr-metric-normalize/
>>
>> Speaking as WG Member:
>>
>> Speaking as WG member, I agree with Peter as well.
>>
>> The additional alternative to local metric normalization is to
do UCMP (Unequal Cost Mulltipath) route installation which is also
a local behavior.
>> Here is one example:
>>
>>
https://www.cisco.com/c/en/us/td/docs/iosxr/ncs5500/routing/79x/b-routing-cg-ncs5500-79x/implementing-ucmp.html
>>
>> You do have assure you don't create loops (e.g., do a feasible
successor check as EIGRP does).
>>
>>
>> Thanks,
>> Acee
>>
>>> On Nov 4, 2025, at 11:46PM, Gyan Mishra wrote:
>>>
>>>
>>> I agree with Peter that this draft is unnecessary.
>>>
>>> Delay metric normalization is common feature implemented by
most all SR vendors to circumvent sub optimal undesired routing
where ECMP exists or TI-LFA backup path is programmed. A single
uS differential in ECMP links can break ECMP as well as TI-LFA.
>>>
>>> This is all local implementation specific and does not require
any standardization.
>>>
>>>
>>> Thanks
>>>
>>> Gyan
>>>
>>>
>>>
>>> On Tue, Nov 4, 2025 at 10:49AM Peter Psenak wrote:
>>> Hi,
>>> here are my comments on the draft:
>>> 1. Normalization of the measured latency, or any other PM
related value, is a local behavior on the router. It should really
be viewed as a part of the measurement procedure. The measured
value is normalized before it is advertised.
>>> 2. It is not functionally required for normalization to be
used, nor for its logic/parameters to be synchronized across all
routers in the network.
>>> 3. Felx-algo definition (FAD) is not meant to be used as a
management mechanism for distribution of the configuration
parameters. There are exiting provisioning tools available for that.
>>> To summarize, I find the proposals in the draft to be both
unnecessary as well as undesirable.
>>> thanks,
>>> Peter
>>>
>>> _______________________________________________
>>> Lsr mailing list -- [email protected]
>>> To unsubscribe send an email to [email protected]
>>> _______________________________________________
>>> Lsr mailing list -- [email protected]
>>> To unsubscribe send an email to [email protected]
>>
>
_______________________________________________
Lsr mailing list -- [email protected]
To unsubscribe send an email to [email protected]
