Louis – I am having increasing difficulty in correlating what you say in this thread and what is actually written in the draft.
Please see responses inline. Look for LES2: From: Louis Chan <lou...@juniper.net> Sent: Tuesday, April 11, 2023 7:46 PM To: Les Ginsberg (ginsberg) <ginsb...@cisco.com>; Acee Lindem <acee.i...@gmail.com> Cc: lsr <lsr@ietf.org>; Krzysztof Szarkowicz <kszarkow...@juniper.net>; Weiqiang Cheng <chengweiqi...@chinamobile.com> Subject: RE: [Lsr] IETF-116 LSR - IGP extensions for Advertising Offset for Flex-Algorithm Hi Les, Thanks for the prompt reply. Please see inline for clarification [lc2]. /Louis From: Les Ginsberg (ginsberg) <ginsb...@cisco.com<mailto:ginsb...@cisco.com>> Sent: Tuesday, April 11, 2023 11:03 PM To: Louis Chan <lou...@juniper.net<mailto:lou...@juniper.net>>; Acee Lindem <acee.i...@gmail.com<mailto:acee.i...@gmail.com>> Cc: lsr <lsr@ietf.org<mailto:lsr@ietf.org>>; Krzysztof Szarkowicz <kszarkow...@juniper.net<mailto:kszarkow...@juniper.net>>; Weiqiang Cheng <chengweiqi...@chinamobile.com<mailto:chengweiqi...@chinamobile.com>> Subject: RE: [Lsr] IETF-116 LSR - IGP extensions for Advertising Offset for Flex-Algorithm [External Email. Be cautious of content] Louis - Please see inline. > -----Original Message----- > From: Louis Chan <lou...@juniper.net<mailto:lou...@juniper.net>> > Sent: Monday, April 10, 2023 11:01 PM > To: Les Ginsberg (ginsberg) <ginsb...@cisco.com<mailto:ginsb...@cisco.com>>; > Acee Lindem > <acee.i...@gmail.com<mailto:acee.i...@gmail.com>> > Cc: lsr <lsr@ietf.org<mailto:lsr@ietf.org>>; Krzysztof Szarkowicz > <kszarkow...@juniper.net<mailto:kszarkow...@juniper.net>>; > Weiqiang Cheng > <chengweiqi...@chinamobile.com<mailto:chengweiqi...@chinamobile.com>> > Subject: RE: [Lsr] IETF-116 LSR - IGP extensions for Advertising Offset for > Flex-Algorithm > > Hi Les, > > Thanks for your questions. Please see inline [lc] below. > > /Louis > > -----Original Message----- > From: Les Ginsberg (ginsberg) <ginsb...@cisco.com<mailto:ginsb...@cisco.com>> > Sent: Saturday, April 8, 2023 7:34 AM > To: Acee Lindem <acee.i...@gmail.com<mailto:acee.i...@gmail.com>>; Louis Chan > <lou...@juniper.net<mailto:lou...@juniper.net>> > Cc: lsr <lsr@ietf.org<mailto:lsr@ietf.org>> > Subject: RE: [Lsr] IETF-116 LSR - IGP extensions for Advertising Offset for > Flex-Algorithm > > [External Email. Be cautious of content] > > > OK - since Acee opened the door - here are some comments from me - > starting with the most important. > > (BTW - I still have limited enthusiasm for this draft.) > > 1)The proposal places some restrictions on how operators provision their > network in terms of assigning SIDs and reserving space for future > assignments. > If operators do not use compatible assignment schemes, then this will never > get deployed. It is therefore not enough to come with a nice idea - you must > have some enthusiasm from the operator community. > > > [lc] If the operator only wants to deploy flex-algo, there is no change in > their > Node-sid numbering scheme. For the Adj-sid, these are local labels with local > significant only, and there is no need for any special planning for Adj-sid, > unless you are suggesting they want to make fixed assignment of Adj-sid > label for each link. Even with fixed, the proposed draft has benefit on that. > I > will explain later. > [LES:] Let's discuss this in the context of prefix-sids - the same applies to adj-sids. Today (i.e., in the absence of your proposal) an operator is free to assign any label within the SRGB for a given prefix/algo pair so long as it is not assigned to some other prefix/algo context. Your proposal places some new restrictions. Now, for a given flex-algo, whenever an operator assigns a given label for a prefix in Algo 0 (call it Label-A0), they must guarantee that "Label-A0+offset" for an advertised flex-algo specific offset is available to be assigned for the prefix/flex-algo pair - and this must be true for all prefixes advertised in algo 0. This is certainly possible to do, but is not guaranteed to be the case in current deployments. For example - and this is only an example...today an operator might utilize a provisioning tool to assign prefix-sids for all supported algorithms on all nodes in the network. To do this, the tool might maintain a database of assigned labels. When provisioning a new node/prefix/algorithm, the logic in the tool might simply take the next available label in the database. The result of this would not be consistent with the requirements of your draft. Which is why I say in order to deploy the extension you propose, such an operator would have to modify its provisioning tool. [lc2] There might be some misunderstanding of our proposal. Let me give some examples. Case 1: Flex-Algo only Prefix offset advertisement: “no” Adj-sid offset advertisement: yes In slide 8’s example, FA129 is using label “402001”, and the advertisement of this label is using existing methods. e.g. SRGB = 400000-460000 FA129: index 2001 (preferred value), or one can choose 111, 222 FA130 (new): index 3001 (preferred value), or 333, 4444 This does not change how the operator to assign label for prefix-sid with their current method. Any index/label could be used for FA prefix within SRGB. The only change is the Adj-sid label allocation, but this “mostly” is only “local” to one node. There is no effect on global label allocation. This draft will be compatible to what operators are doing today. Case 2: VFA only Prefix offset advertisement: yes Adj-sid offset advertisement: yes I agree, with VFA, there would be impact to global allocation to node-sid/prefix-sid. But VFA is a totally new concept. No one has deployed that yet. There is no impact to operators which stick to deploy only Flex-algo. Other case: Flex-Algo w/o Adj-sid offset Continue the example of Case#1 above Another FA131 is added, but no Adj-sid offset is advertised The question would be * Either allow this configuration, and FA131 will fallback using Algo 0’s adj-sid * Or, disallow this configuration I tend to “allow” such configuration with mix of FA129, FA130 (with adj-sid offset) and FA131 (w/o adj-sid offset) [/lc2] [LES2:] Section 4 of the draft defines three new sub-TLVs for advertising offset. One for adj-sid. One for LAN adj-sid One for prefix sid. They all have the following fields/description: Algorithm: 1 octet Flex-Algo value is between 128 and 255 Virtual Flex-Algorithm 4 octets Value 0 means that it is not VFA Value >=256 means VFA identification number, and the base topology is denoted by Algorithm Value 128 to 255 are invalid Yet, you say here that the new prefix-SID sub-TLV is not meant to be used for flex-algo – only your new construct VFA. How would I draw that conclusion from the text in the draft??? But, let’s look at adj-sid since the same issues arise – though they are a bit more complicated in the adj-sid case (which is why I started with prefix-sid). Adj-sids have a number of attributes – one of which is “persistence” (the P-flag in https://www.rfc-editor.org/rfc/rfc8667.html#section-2.2.1 – which you do not model in your draft). Persistent adj-sids are meant to have the same value across node reboots – and are therefore assigned in a manner similar to prefix-sids. Typically they would be allocated from the SRLB on the advertising node. In which case the same set of issues I discussed in the context of prefix-sids applies. For non-persistent adj-sids, these could take on different values when a node is rebooted – or even when an interface flaps. Typically, they would be allocated from an unreserved pool of labels. With your proposal, for each interface, whenever an adj-sid is assigned, additional adj-sids have to be reserved for each supported flex-algo. To use a simple example, consider that we support only one flex algo (128) and the offset for this algo is 1000. Interface 1: Base label: 1000 Flex label 2000 Interface 2: Base label: 1001 Flex label 2001 Etc. If I initially have 500 interfaces in use by IS-IS, I have now assigned/reserved labels 1000-1499 and 2000-2499. If I subsequently enable 100 additional interfaces for IS-IS, I would need to obtain additional pairs of labels, but as I am allocating these labels out of a space reserved for dynamic allocations, it isn’t guaranteed that Label N and Label N+1000 are both available. And since non-persistent adj-sids can be assigned/deassigned dynamically, over time the unreserved pool of labels may get increasingly fragmented, making it more difficult to obtain the necessary pair of labels needed. Add to this the difficulties of introducing support for an additional flex-algo on the fly and the need to support both “protected” and “unprotected” adj-sids (the B-flag) and I think implementations would find it quite challenging to guarantee they can always obtain the set of labels required by offset advertisement. Also you require implementations to support atomic allocations of a group of labels – something which I would not assume is easily supported. Les Could this be done? Sure. Do operators want to do this? I do not know. But since this would be necessary in order to use your proposed extension, it is necessary to gauge operator enthusiasm for making such changes in order to know whether there is any point in proceeding with your proposal. > In slide 8 of the below presentation > https://datatracker.ietf.org/meeting/116/materials/slides-116-lsr-03-ietf116-<https://urldefense.com/v3/__https:/datatracker.ietf.org/meeting/116/materials/slides-116-lsr-03-ietf116-igp-adv-offset-01__;!!NEt6yMaO-gk!Bl7Swe9ql9VT0qGkD6FoZZzTWT2fmIx55eSncdmMgoCJetJ5-80micuqnqk79yewGB-BleOfrYpSjfI$> > igp-adv-offset-01<https://urldefense.com/v3/__https:/datatracker.ietf.org/meeting/116/materials/slides-116-lsr-03-ietf116-igp-adv-offset-01__;!!NEt6yMaO-gk!Bl7Swe9ql9VT0qGkD6FoZZzTWT2fmIx55eSncdmMgoCJetJ5-80micuqnqk79yewGB-BleOfrYpSjfI$> > > FA129 is a prefix-sid (400201) allocated by operator, and it can be any label. > There is no connection to how Adj-sid is derived. > Per Flex-Algo adj-sid assignment is not affecting network wide label > assignment from operation perspective. Each node could have different local > block for such adj-sid assignment. One might need to estimate total possible > number of link in one chassis for such allocation, or it could be estimated by > OS software itself. I also mentioned in the session, if there is 100 x FA with > 1000 links (high end), it is only 100k labels. Is it difficult to allocate > such. [LES:] Your proposal does not reduce the number of labels which need to be "allocated" and installed in forwarding, it only reduces the number of bytes used to advertise this information in LSPs/LSAs. [lc2] You are correct. I think, at the same time, it helps reducing time for global convergence since the advertisement size is smaller, especially in network with long diameter with multi-hops. Also, in https://datatracker.ietf.org/doc/draft-ietf-lsr-isis-fast-flooding/ (which you participated in) >>> As IS-IS is deployed in greater scale both in the number of nodes in an area and in the number of neighbors per node, the impact of the historic flooding rates becomes more significant. Consider the bringup or failure of a node with 1000 neighbors. This will result in a minimum of 1000 LSP updates. At typical LSP flooding rates used today (33 LSPs/second), it would take 30+ seconds simply to send the updated LSPs to a given neighbor. Depending on the diameter of the network, achieving a consistent LSDB on all nodes in the network could easily take a minute or more. <<< This proposed draft will certainly help. [/lc2] > > So, this is why I do not understand your question in full. > > If the operator plans to use VFA, that would be a different discussion. VFA, > today, does not exist in deployment. > [LES:] My comments here have nothing to do with VFA. > [/lc] > > Have you discussed this idea with any operators? > > [lc] I include Wei-qiang from China mobile in this thread. He has shown his > need on this kind of solution. Maybe, he could give his perspective here. > [/lc] > > If so, what has been their response? > If they are open to the idea, how might they migrate from their existing > assignment schemes to an assignment scheme compatible with the > proposal? > > These are questions that need to be answered before considering this idea. > > [lc] In slide 8, if you see these label numbers > > Prefix-sid: 400001, 402001, 406001, 407001 > Adj-sid: 32, 2032, 6032, 7032 > > From operator perspective or troubleshooting perspective, value xxx001 > represent the same node, and value x032 represent the same link. This > makes things more organized and easier to understand. > > If all are random labels, I do not see any benefit at all. > [/lc] [LES:] I am not commenting on whether the label assignment scheme you propose is better or worse than any other. I am only pointing out that you are imposing new restrictions on how labels are allocated. As you are not in charge of how operators provision their networks (nor am I), it is presumptuous of you to think that simply because you think this is a better way to do things that operators will be happy to modify their existing networks to conform to your proposed restrictions. This isn’t academia - you need to vet this with the operator community. [lc2] Please refer to the examples at the top. The picture should be clear by now. There is no restriction to what is deployed today. [/lc2] > > 2)Section 5 Compatibilty > > There is no "compatibility" with legacy nodes - because all nodes in the > network have to have a consistent understanding of what SID is assigned to a > given context (for prefixes and adjacencies) since they might need to install > forwarding entries for that context. > I do not see any point in deploying this until all nodes support it. If you > did do > so, you would need to advertise old and new forms - which does the > opposite of what you are trying to achieve. Instead of reducing LSP space > used you would increase it. > > [lc] If the operator just plans to use only Flex-Algo, and no VFA, it should > be > compatible with legacy implementation. If legacy nodes do not understand > adj-sid offset notation, these nodes could just ignore it. The forwarding > plane should work with co-existence of old and new nodes. Per flex-algo adj- > sid is only local significant to one node. New nodes should detect whether > legacy nodes exist in the network via such new extension advertisement. > And new nodes should use only algo 0 adj-sid from legacy nodes for any TI- > LFA. [LES:] Consider a network of 100 nodes. Let's say the "left-hand-side" of the network consist of legacy nodes who do not understand your new advertisements. The "right-hand-side" of the network consists of upgraded nodes who support the new advertisements. Consider nodes PE-LEFT and PE-RIGHT. PE-RIGHT advertises a prefix-SID of 1000 for 2.2.2.2/32, and an offset of 1000 for flex-algo 128. PE-LEFT supports flex-algo 128 and wants to install a forwarding entry for 2.2.2.2 for flex-algo 128. It looks in the LSPs originated by PE-RIGHT. It does not see any assigned SID for 2.2.2.2/32 flex-algo 128. It cannot create a forwarding entry. And neither can any other node in the left hand side of the network. When PE-RIGHT stops advertising the explicit prefix SID for 2.2.2.2/32 Algo 128, all legacy nodes are unable to create forwarding entries for the prefix/algo tuple. This isn’t backwards compatible. 😊 In general, you cannot advertise information legacy nodes require in a new container that legacy nodes do not understand and claim that you are backwards compatible. [lc2] Please refers to the examples for clarification. 1. For existing Flex-Algo deployment, it would be compatible. There is no change in the container format on how prefix-sid 2.2.2.2/32 in FA128 is advertised. 1. For new VFA, it would not be compatible. But….it does not mean that we could not have VFA running in the same network. There could be procedures to enhance such. With your example, one workaround could be: For VFA 600, PE-RIGHT detects that PE-LEFT does not participate in VFA600 (due to no offset advertisement seen), * Either, it spawns new CSPF for VFA600 instead of sharing FA129’s result. Bypass PE-LEFT as a result. * Or, it uses legacy node FA129 prefix-sid and adj-sid as replacement (note: this method needs more comment) In either ways, VFA600 could work without issue even with legacy nodes co-existence. After PE-LEFT upgraded, VFA600 would be using FA129 CSPF result instead, and save CPU resources in each node. Another question: do we need FAD for VFA600? Currently, no. Not mandatory. But it could be considered if “good to have” parameters are passed along with FAD. [/lc2] > > I do not see a major problem. Please give me an example to illustrate your > concern if possible. > > Of course, we need to do double check on the claim and possibly lab > verification to see if the backward compatibility could be achieved. It could > be vendor specific. > > [/lc] > > > What does deserve discussion is a "hitless migration strategy". When full > support is available, if you were to switch to the new scheme, you would > want to do so without changing any existing SIDs as this would avoid > forwarding disruption. Which means operators would have to modify their > SID assignment scheme in advance of deploying the new scheme. > > [lc] For VFA, there would be issue for legacy nodes. I agree. In this case, > solution could be > - either have a fallback plan for newer nodes if detection of legacy > nodes > exist in the network. E.g. spawn new CSPF > - or, totally not to deploy VFA unless all nodes are upgraded. > > Section 5 is not updated with VFA inclusion. [LES:] My comments have nothing to do with VFA. Please reconsider them after you have understood the backwards compatibility issues. > > [/lc] > > 3)Virtual Flex Algorithm > > You have introduced a new concept with very little explanation of what it is > nor how it can be supported. > For example, how would we determine which nodes support a given VFA? > Since the algorithm value must be greater than 255, it cannot be advertised in > the existing SR Algorithm sub-TLV. > > If you are serious about this idea, please provide a more complete > discussion. > > [lc] We could illustrate application examples in next presentation. For > ethernet, we have port, and then we have VLAN and stacked vlan. History > has some hints on this. > [LES:] You are writing a normative specification. Hoping that all readers/implementors have the same "intuition" isn’t sufficient. Les > [/lc] > > 4)Section 4.3 > > "R" and "N" flags are now defined in prefix attributes sub-TLV (RFC7794) > They were originally defined in the SR sub-TLV because RFC 7794 did not exist > at the time. > The only reason they continue to exist in RFC 8667 is for backwards > compatibility with early implementation of SR-MPLS based on early drafts of > what became RFC 8667. > Please do not introduce them in new sub-TLVs - there is no need. > > [lc] noted with thanks [/lc] > > 5)ADJ-SIDs are NOT allocated from the SRGB as they are local in scope. > They MAY be allocated from the SRLB - or outside either GB range. > Please correct the document in this regard. > > [lc] noted [/lc] > > Les > > > -----Original Message----- > > From: Lsr <lsr-boun...@ietf.org<mailto:lsr-boun...@ietf.org>> On Behalf Of > > Acee Lindem > > Sent: Friday, April 7, 2023 11:43 AM > > To: Louis Chan <lou...@juniper.net<mailto:lou...@juniper.net>> > > Cc: lsr <lsr@ietf.org<mailto:lsr@ietf.org>> > > Subject: Re: [Lsr] IETF-116 LSR - IGP extensions for Advertising > > Offset for Flex-Algorithm > > > > Hi Louis, > > > > In the interest of initiating discussion, I would like to propose the > > term "Flex Algorithm Traffic Class (FATC)" for the sub-division of > > flex-algorithm traffic referred to in the draft as “Virtual Flex Algorithm”. > > > > Also, in your terminology, you refer referred to TLVs and fields with > > simply “Algorithm” when RFC 9350 uses “Flex Algorithm”. > > > > Thanks, > > Acee > > > > > > > > _______________________________________________ > > Lsr mailing list > > Lsr@ietf.org<mailto:Lsr@ietf.org> > > https://urldefense.com/v3/__https://www.ietf.org/mailman/listinfo/lsr_<https://urldefense.com/v3/__https:/www.ietf.org/mailman/listinfo/lsr_> > > _;!!NEt6yMaO-gk!B9ufrV6k- > c7mtP9JUiXbrF3NCkZ15_UMLBjV_fnJovfz18M5VkkI2F > > EoixpkxsfMnbqwbR0bpHgoS9E$ > > Juniper Business Use Only Juniper Business Use Only Juniper Business Use Only
_______________________________________________ Lsr mailing list Lsr@ietf.org https://www.ietf.org/mailman/listinfo/lsr
