Re: [Teas] Default NRP definition [Was: Repeated call for last call on draft-ietf-teas-ietf-network-slices]

[Krzysztof Szarkowicz <[e-mail]>]

Adrian,

I have couple of questions here:


1. Taking into consideration typical SP network today, where we have:

a) differentiated services realized via mapping of DCSP and/or MPLS TC values to buffers, and deploying some differentiated scheduling
b) running services (L3VPN, L2VPN, ...) over such network
c) possibly (but not necessarily) deploying some TE

Do we refere to typical current SP deployment as using ’single NRP’ or not using NRP at all?


2. If I have in my network two set of tunnels between PE nodes, using different link metric types (e.g. one set of tunnels uses IGP link metric to determine the path through the network, another set of tunnels using TE link metric to determine the path through the network), and these two sets of tunnels use exactly the same resources: entire topology, i.e. all links and nodes in the network, and the PHB is exactly the same (i.e., packet with QoS marking ‘X’ get exactly the same treatment in terms of buffering/scheduling, regardless if forwarded over tunnel from 1st tunnel set, or tunnel from 2nd tunnel set) are we talking about one NRP or two NRPs?

//Krzysztof

> On 2022 -Sep-22, at 18:30, Adrian Farrel <[e-mail] <mailto:[e-mail]>> wrote:
> 
> John makes some good points.
>  
> Adding a definition of a term that is only used in parentheses in one (early) individual draft where one of the authors says it was a mistake to use it, seems excessive. Perhaps we should all just stop using the term?
> 
> The idea of “default” seems wrong in any case. NRPs should be explicit. Sure, you can have a single one that includes all resources on all links, but that is still an active choice.
>  
> Adrian
>  
> From: Teas <[e-mail] <mailto:[e-mail]>> On Behalf Of John E Drake
> Sent: 22 September 2022 14:55
> To: John E Drake <jdrake=[e-mail] <mailto:jdrake=[e-mail]>>; [e-mail] <mailto:[e-mail]>; [e-mail] <mailto:[e-mail]>
> Subject: Re: [Teas] Default NRP definition [Was: Repeated call for last call on draft-ietf-teas-ietf-network-slices]
>  
> Adrian,
>  
> Upon reflection, the revised wording changes the meaning.  We start by observing that “The connected set of links can be the entire set of links in the underlay network” and then continue with “ *and in this case there
> can be a single NRP* and it has all of the buffer/queuing/scheduling resources for each of the links in the underlay network”.  I.e.,  We can define one or more NRPs that use the entire underlay network topology but we can also define, in this case, a single NRP that uses all of the underlay network resources – the underlay network has a topology and it has resources. 
>  
> Yours Irrespectively,
>  
> John
>  
>  
> Juniper Business Use Only
> From: Teas <[e-mail] <mailto:[e-mail]>> On Behalf Of John E Drake
> Sent: Thursday, September 22, 2022 9:01 AM
> To: [e-mail] <mailto:[e-mail]>; [e-mail] <mailto:[e-mail]>
> Subject: Re: [Teas] Default NRP definition [Was: Repeated call for last call on draft-ietf-teas-ietf-network-slices]
>  
> [External Email. Be cautious of content]
>  
> Adrian,
>  
> I am okay with your revised wording for single NRP, but I don’t agree that we need to define a ‘default NRP’ because it is attempting to detail how a given service provider *might* operate its underlay network.  I.e., it is pure speculation.
>  
> Yours Irrespectively,
>  
> John
>  
>  
> Juniper Business Use Only
> From: Teas <[e-mail] <mailto:[e-mail]>> On Behalf Of Adrian Farrel
> Sent: Thursday, September 22, 2022 6:19 AM
> To: [e-mail] <mailto:[e-mail]>
> Subject: [Teas] Default NRP definition [Was: Repeated call for last call on draft-ietf-teas-ietf-network-slices]
>  
> [External Email. Be cautious of content]
>  
> Hi all, again.
>  
> Jumping in at the top of the thread, yet again, to try to dig into two pieces of terminology. Picking up particularly on Greg, Jie, and Pavan’s points.
>  
> “Single” does, indeed, mean “just one”. But it’s usage is very deterministic, meaning “one of (potentially) many” in some cases, and meaning “there is exactly one” in other cases. Perhaps it would help if:
> OLD
>    The connected set of links can be the
>    entire set of links in the underlay network and in this case there
>    can be a single NRP and it has all of the buffer/queuing/scheduling
>    resources for each of the links in the underlay network.
> NEW
>    The connected set of links can be the
>    entire set of links in the underlay network and in this case there
>    can be precisely one NRP supported in the underlay network where
>    that NRP has all of the buffer/queuing/scheduling resources for
>    each of the links in the underlay network.
> END
>  
> “Default” has, of course, a clear meaning in English (although there are several different meanings). As engineers, we should be careful not to introduce terms without also writing a clear definition. If we want to use the term “default NRP” then we should define it and, in that case, this document seems like a fine place to include it. But we are definitely fishing around for what “we” mean by the term. I think we are getting to…
>  
> Default NRP:
>    The default NRP is constructed from all of the buffer/queuing/scheduling
>    resources on all of the links in the underlay network that have not been
>    assigned for use by any other NRP.  That is, it consists of the residue 
>    resources.  If no other NRP has been defined, the default NRP comprises
>    all of the buffer/queuing/scheduling resources of the underlay network.
>    If a further NRP is subsequently defined, the default NRP will be reduced
>    by the resources assigned to the new NRP.  If an NRP is deleted, its
>    resources are released back into the default NRP.
>  
> Commensurate with that, the text quoted above could can become…
>    In the case where there is just the default NRP and no other NRPs
>    have been defined, the connected set of links can be the entire set
>    of links in the underlay network, and in this case there is precisely
>    one NRP (the default NRP) supported in the underlay network where
>    that NRP has all of the buffer/queuing/scheduling resources for
>    each of the links in the underlay network.
>  
> Thoughts?
>  
> Cheers,
> Adrian
>  
> From: Vishnu Pavan Beeram <[e-mail] <mailto:[e-mail]>> 
> Sent: 22 September 2022 06:34
> To: [e-mail] <mailto:[e-mail]>
> Cc: Krzysztof Szarkowicz <[e-mail] <mailto:[e-mail]>>; Joel Halpern <[e-mail] <mailto:[e-mail]>>; [e-mail] <mailto:[e-mail]>; John E Drake <jdrake=[e-mail] <mailto:jdrake=[e-mail]>>
> Subject: Re: [Teas] Repeated call for last call on draft-ietf-teas-ietf-network-slices
>  
> Adrian, Hi!
>  
> Thanks for the top-post. Please see inline (prefixed VPB).
>  
> Regards,
> -Pavan
>  
>  
> On Thu, Sep 22, 2022 at 3:46 AM Adrian Farrel <[e-mail] <mailto:[e-mail]>> wrote:
>> Hi,
>> 
>> Sort of top-posting on the thread, and speaking as editor.
>> 
>> Krzysztof >>
>> > I see that the current text is clear and precisely describes the
>> > intent of single (default) NRP, so it doesn’t need any change/correction.
>> 
>> Well, it was certainly the intent that the text would be clear, but if some people are confused or unclear, we should seek to make things clearer.
>> 
>> Note well that the term "default NRP" is not one that is used in the document, and any lack of clarity about the term must be laid at the feet of the people using the term!
>> I *think* the term is being used to describe the limiting case where there is just one NRP that is all of the resources in the network.
>> 
>> Joel >>
>>  > Does that single NRP admit multiple diffserv code points / queueing behaviors?
>> [JD]  That is at the discretion of the underlay network operator 
>> 
>> I think John and Joel may be at cross-purposes with the same conclusion.
>> To Joel: Yes, the single NRP admits the possibility of multiple diffserv code points / queueing behaviors.
>> To John: Yes, the underlay network operator is free to make the default NRP have multiple or fewer codepoints / queueing behaviors.
>> 
>> Joel >>
>> > If so, then the notion of NRP is itself purely an arbitrary collection of
>> > behaviors, and thus not helpful or particularly meaningful. 
>> 
>> "Arbitrary" and "helpful" are possibly a bit loaded.
>> Recall that the NRP is an internal mechanism for the underlay network operator. It is not exposed to the customer, but is a tool for the operator.
>> It allows the operator to partition their network in a way that they find useful for the rapid construction of network slices. 
>> What that amounts to is that the operator may profile the resources of the network into collections (NRPs) to enable the support of particular types of network slice service.
>> The way that an operator does this is entirely up to them (it's a policy), so it could be arbitrary or highly logical.
>> 
>> But some people think that it won't be necessary to build NRPs and so we have the concept of "the default NRP" which is essentially all of the resources of the network.
>> It's a null-op in the process, but we keep it there to have a consistent picture.
>> 
>> Joel >>
>> > One way out is to declare that relative to any given device, the collection of behaviors in
>> > an NRP may be different diffserv code points but may not be further differentiated.  
>> > Another way out is to declare that the collection referred to in the definition refers to
>> > the collection across devices, but within a device an NRP has only one queueing
>> > behavior / resource.
>> 
>> But I wonder if there is a confusion between resources and behaviors? The text in the draft is clear that it is describing resources. How the resources are used is surely a different matter, or is it?
>> 
>> As a quick reference, the text we're talking about is...
>> 
>>    A Network Resource Partition (NRP) is a collection of resources
>>    (bufferage, queuing, scheduling, etc.) in the underlay network.  The
>>    amount and granularity of resources allocated in an NRP is flexible
>>    and depends on the operator's policy.  Some NRP realizations may
>>    build NRPs with dedicated topologies, while some other realizations
>>    may use a shared topology for multiple NRPs; one possible realization
>>    is of a single NRP using all of the resources of the entire underlay
>>    network topology.  Thus, an NRP consists of a subset of the
>>    buffer/queuing/scheduling resources on each of a connected set of
>>    links in the underlay network.  The connected set of links can be the
>>    entire set of links in the underlay network and in this case there
>>    can be a single NRP and it has all of the buffer/queuing/scheduling
>>    resources for each of the links in the underlay network.
>> 
>> Pavan and Lou >>
>> > This thread does seem to suggest there are some loose ends with 
>> > respect to the notion of a default NRP that need to be tied before
>> > publication. There are some open questions on how resources in 
>> > the default NRP get impacted when you start adding resource
>> > partitions in the underlay network. 
>> 
>> We do have to return to ask, "What is this default NRP that you are talking about?" If it is, as I assume, the "single NRP" that "has all of the buffer/queuing/scheduling resources for each of the links in the underlay network" then it should be fairly obvious that adding other NRPs does change the definition of the "default NRP." This happens because the default NRP stops being the only NRP and so stops being the default NRP.
>> 
>> I believe you have yourself wrapped around the definition of a term that doesn't exist.
>  
> [VPB] You are right -- draft-ietf-teas-ietf-network-slices does not use the term "default NRP".  draft-ietf-teas-ns-ip-mpls, which extensively discusses the notion of one or more network resource partitions, also does not use this term (yet). But, we are starting to discuss slicing realization documents in the WG that are building on this notion of a "default/single/only NRP" as framed in draft-ietf-teas-ietf-network-slices (see draft-srld-teas-5g-slicing which does use this term) and for that purpose it may be useful to discuss what this entails (now rather than later).  As Jie has pointed out in this thread, there is an interpretation here that you may start with a default NRP (no explicit resource partitioning) to realize slicing, but you may end up having the default NRP co-exist with non-default NRPs as they get gradually added to the network. The default NRP in this interpretation may simply translate to the set of resources that don't meet the selection criteria of any explicit user-specified NRP (if there are no user-specified NRPs, then the default NRP includes all the resources in the underlay network). Another interpretation of the default NRP is (like you said) that it ceases to exist when the first resource partition is made (two explicit NRPs get created).
>  
> [VPB] We (the WG) may end up saying that we don't need to discuss "default NRP" or its semantics in draft-ietf-teas-ietf-network-slices, but rather have it discussed in draft-ietf-teas-ns-ip-mpls (which does talk about slicing realization using one or more resource partitions) instead. But it is a loose end that needs to be tied at some point. 
>  
>> 
>> Pavan and Lou >>
>> > We are hoping that the WGLC (the process for which has just begun)
>> > would be a forcing function for those of you (chairs included) who
>> > intend to suggest text/edits to clear this up.
>> 
>> It would be great if exactly that happened. That is, text suggestions.
>> 
>> Cheers,
>> Adrian 
> _______________________________________________
> Teas mailing list
> [e-mail] <mailto:[e-mail]>
> https://www.ietf.org/mailman/listinfo/teas <https://www.ietf.org/mailman/listinfo/teas>