[mpls] Re: [spring] Hop by Hop TE in SR networks (aka draft-beeram-mpls-rsvp-sr-01)
Robert Raszuk <robert@raszuk.net> Thu, 12 November 2026 17:27 UTC
Received: from mail-qv1-xf2c.google.com (mail-qv1-xf2c.google.com [IPv6:2607:f8b0:4864:20::f2c]) (using TLSv1.3 with cipher TLS_AES_256_GCM_SHA384 (256/256 bits) key-exchange x25519 server-signature ECDSA (prime256v1) server-digest SHA256) (No client certificate requested) by mx.ietf.org (Postfix) with ESMTPS id CB25F46 for <mpls@ietf.org>; Thu, 12 Nov 2026 17:27:07 +0000 (UTC)
Authentication-Results: mx.ietf.org; dkim=pass header.d=raszuk.net header.s=google header.b=GDOp5kuL; spf=pass (mx.ietf.org: domain of robert@raszuk.net designates 2607:f8b0:4864:20::f2c as permitted sender) smtp.mailfrom=robert@raszuk.net; dmarc=none; arc=pass ("google.com:s=arc-20260327:i=1")
Received: by mail-qv1-xf2c.google.com with SMTP id 6a1803df08f44-91f1e3dd08cso188126d6.3 for <mpls@ietf.org>; Thu, 12 Nov 2026 09:27:07 -0800 (PST)
ARC-Seal: i=1; a=rsa-sha256; t=1794504427; cv=none; d=google.com; s=arc-20260327; b=ArT1ebXP3bMKaKRH6VeDxWFLg6SbcGcMrb0Au227EvVPPQocbTP4kEEh1yu6O97W9i qUweaCn03v5JIYhARZbji2BJXjUG7f1iA2GrdAzeiDatunNOm3K+TcZSYsFuHyZorfBy EO34v4rOTPvvm4OR066YBaNN4utlN4yFpoGY4dhI7U18lMdsMpzYo7bMUjSxYfhVt/js KVXDQeDTKJIk6O/0Iua0Kge65RZyl8hG6mn27h7hepSkqXxeOmWqaMxhE5puHUr1+YsP ksjIxAyPli6TUaLUjERNGkE1JrrI6/2oB7QLSPcihUAK1UQXik7CKNp7le0lBienYd51 wNkg==
ARC-Message-Signature: i=1; a=rsa-sha256; c=relaxed/relaxed; d=google.com; s=arc-20260327; h=cc:to:subject:message-id:date:from:in-reply-to:references :mime-version:dkim-signature; bh=IhGfkx3eN+R6vBKQWJ8vn4H1+1t8pdDfRX7DNdovGkY=; fh=lNZrPDpZpi8kMjbH3JuACllcfY9Gfo//c0uTb4TYXmE=; b=Ep68fpcbn8ERzMZnzvCzN0Cb2dWXkZE3emOUYwJKx3psI6gdQFSfqUXvRDZT63kbLL xkw8zALvFxyWTKpQzVd/HWmv1oJlzjRILAbOOGsye2XkajVMgL8fxRpVyX//H0fqVxYh WQFAg3sR6OZVHuToot8e6PMEhxpwpQEag619XCbyfvaUs8brovl2m70cMNSM2dguz2Pp Cgf6LL/r6fDoBYtwD3mYvukWq5rQsglhbJZW9fK09Zckri6Wgrw+m8EH/GoCfLLCqUAZ ep2JmHFl8lfEs0bOCDhsy2UcCGlWXx0pqvPczY7Vlh11oM8UW+36eeCWy5elaz1A8FOA ow9A==; darn=ietf.org
ARC-Authentication-Results: i=1; mx.google.com; arc=none
DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=raszuk.net; s=google; t=1794504427; x=1795109227; darn=ietf.org; h=content-type:cc:to:subject:message-id:date:from:in-reply-to :references:mime-version:from:to:cc:subject:date:message-id:reply-to :content-type; bh=IhGfkx3eN+R6vBKQWJ8vn4H1+1t8pdDfRX7DNdovGkY=; b=GDOp5kuL8bHB/AH+hTE4dQRs2GMxLXypcMbzNRhNuNFHpokod4DWgBYktH3iPeunf5 HdYoy4YHKjMWsFX2EZQqqNj74tPkNXT8xdpAepUsH3jZhCXVVAJ6+7n3aD9wbrWSMbyf DCCnWdC0p0FLAiOEag43F7YSG7pYUhWevsppbuoGCu5TGkg5A9zL0aF3lKi+2YhyoKOH L3qlLCYng19TSKLy5/+YeRc1VL6G1d6zc/oGLQcEiPCCYzPM+uTBKH1pFmFs8fTmC8s5 /gH9rKItsxnwHcky3PWOkckQC6ZS6GXsfUx9ntTrWOP76NTZeDwrM7nkW+jCnDq5eANI fjVg==
X-Google-DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=1e100.net; s=20260707; t=1794504427; x=1795109227; h=content-type:cc:to:subject:message-id:date:from:in-reply-to :references:mime-version:x-gm-gg:x-gm-message-state:from:to:cc :subject:date:message-id:reply-to:content-type; bh=IhGfkx3eN+R6vBKQWJ8vn4H1+1t8pdDfRX7DNdovGkY=; b=AWsgd1SvBXUTf41F3Ow7SOH/JAnbLkIaR0+lXDeHcyqF0kr2PJcdmtgjH+ShqAGrVE jIJ2oqD6UWmUUzrnqIRHq/sJeh0sSxnLoqtLs3TJ+HNXPwDAIlaRKJqASXlEVHaDhUGp m5tSs38vBHg20ojsUFsUuF3uS0wYClXYOL4oAwMcWjxwpBJvvnDI/2JfwwZXW1auEiBZ NKeexjC04EnQ7GPzPiQAjdGnZ1Zf822mSv7yqMJA8RxQnPZnzTJDcYmag7lTr3TKnWaX nAbNZUuji2HHs3yGKITk1I+B4PEt4xxQq9fjyYWmpH4Q1cwJeq1GBE4ngTVAjhe5iA0w 2CCg==
X-Forwarded-Encrypted: i=1; AL+hErFma/tH7P5e6RkOKbY/kzTdiXbEjnJGXs3Eh3OrsC5IhJ25Ii3OJH0GYPUdPiIFD8afEfmq@ietf.org
X-Gm-Message-State: APuU4Z/74O2JCdmW3VYiOxS6QJJWHnrcLz8OniF9hbjmm1wbuqv+xP7V w3OqowVcJYUSU7gEcymV8IKKRkDhiTHxVDyKxuxiN3EhKQOUtpQjCyqTcnwWDUTUDZkkoian4KV FZjRPSYWkxheAZlfhPc/cFmbHsckWe1ESuOcsNahxYg==
X-Gm-Gg: AYBFou1PacIILw+RNwv3V0Iv0XmrIQacEFXPX+ZYjZb3azsCyJ+xKV7q6KeFiJTmXFW ZDwdm4WmjJdZ84yxnAmJBTZvdkfUfoqB/4yCXbFGqteQhunpsG8Z2IWlhObH7rdG1dnVefPzr3X RlaQ0YgIr6nYrTvQxhDsB9C9xcE1fd8AdGCbUrXqB7JfuKxj8UyXJKZ4uhGjuOJ9Y1z40NhwJlm LWISmJjKMjpoUIUTdtpPjMq056ksWFur9HFnBkqxoju1C7vcUzKZc9mGIJpFYAt2AxoicCClqUz MZg3CjnJRCO6NJTDgHlICciA8T85DGovrgjNHfM4YQkaSvWE2u4bHsCYHrtPW+Wu23xzXCBbwUn PeZauqpLCt5y7ShDiYE2j4MKZIMIT6c6jIbMUqw==
X-Received: by 2002:a05:6214:4886:b0:91e:b7fd:b35f with SMTP id 6a1803df08f44-91f0409994cmr103205136d6.30.1794504426917; Thu, 12 Nov 2026 09:27:06 -0800 (PST)
MIME-Version: 1.0
References: <CAOj+MMGNFOvZvsyoMU1oDi78PUs_xfe2aSNJWpE6uPRvjCGzsA@mail.gmail.com> <F55A428D-47B4-4EC4-A184-0CA2733F051C@arrcus.com> <SJ2PR84MB3370BA05E77F332076730FBD92702@SJ2PR84MB3370.NAMPRD84.PROD.OUTLOOK.COM>
In-Reply-To: <SJ2PR84MB3370BA05E77F332076730FBD92702@SJ2PR84MB3370.NAMPRD84.PROD.OUTLOOK.COM>
From: Robert Raszuk <robert@raszuk.net>
Date: Thu, 12 Nov 2026 18:26:56 +0100
X-Gm-Features: AclHuK_-OCx5X73FCeqFy3iKyVc7xuB7NlEw_cxkI5nGdeOXxPWl7Zmb0sjJINg
Message-ID: <CAOj+MMGKSdDF1Q+wf9Z7ctUe5hGwKVCiXjBf++Uy1ZJBGboeSw@mail.gmail.com>
To: "Barth, Colby" <jonathan.barth@hpe.com>
Content-Type: multipart/alternative; boundary="000000000000ea23480660179e90"
X-Spamd-Bar: ----------------
Message-ID-Hash: Y4RZFFUT3RPW3RPKXN3SISSI4VA233GE
X-Message-ID-Hash: Y4RZFFUT3RPW3RPKXN3SISSI4VA233GE
X-MailFrom: robert@raszuk.net
X-Mailman-Rule-Misses: dmarc-mitigation; no-senders; approved; loop; banned-address; header-match-mpls.ietf.org-0; emergency; member-moderation; nonmember-moderation; administrivia; implicit-dest; max-recipients; max-size; news-moderation; no-subject; digests; suspicious-header
CC: Andy Smith <andy@arrcus.com>, mpls <mpls@ietf.org>, TEAS WG <teas@ietf.org>, SPRING WG <spring@ietf.org>, "Andrew Stone (Nokia)" <andrew.stone@nokia.com>
X-Mailman-Version: 3.3.10
Precedence: list
Subject: [mpls] Re: [spring] Hop by Hop TE in SR networks (aka draft-beeram-mpls-rsvp-sr-01)
List-Id: Multi-Protocol Label Switching WG <mpls.ietf.org>
Archived-At: <https://mailarchive.ietf.org/arch/msg/mpls/Q2-cyOIFw2ksYP5J1BYT2PtVfK8>
List-Archive: <https://mailarchive.ietf.org/arch/browse/mpls>
List-Help: <mailto:mpls-request@ietf.org?subject=help>
List-Owner: <mailto:mpls-owner@ietf.org>
List-Post: <mailto:mpls@ietf.org>
List-Subscribe: <mailto:mpls-join@ietf.org>
List-Unsubscribe: <mailto:mpls-leave@ietf.org>
Hi Colby, Thx for the tutorial on auto-bw. Just FYI ... I delivered one too in Feb 2000 https://nanog.org/events/nanog-18/content/3413/ Then was for a number of years heavily involved in various MPLS TE global deployments. > Second, many modern bandwidth-engineered TE networks have paths exceeding 15 hops. Good luck with putting SRH and 15+ SIDs not to count SIDs for actual applications. Sure uSIDs may help you significantly here. But let's be very honest ... SRv6 encap with say 20 SIDs is a much bigger burden to data plane then same encap with 2-3 SIDs. Yes one can say this is stateless in data plane and only stateful in control plane ... but packet size and packet processing does matter. > Third, failures are where bandwidth management really matters. Absolutely agree. And first the story you are painting works only when you have 100% TE in the network not a mix of TE and non TE traffic. Then computing bypass with BW awareness in a distributed fashion is IMO a big challenge. All the best, Robert On Thu, Nov 12, 2026 at 5:57 PM Barth, Colby <jonathan.barth@hpe.com> wrote: > Hi Robert, > > To add a little to Andy’s response. > > A couple of observations from my experience with bandwidth-engineered TE > networks: > > First, auto-bandwidth addresses the question of how much bandwidth to > reserve. It's widely deployed and has proven effective in hundreds of large > TE networks. The concept is simple; Traffic statistics are collected at the > ingress, per TE tunnel, and used to adjust bandwidth reservations as demand > changes. CSPF uses these updated bandwidth requirements, along with > available network resources and TE constraints, to find “better" paths, > avoid congestion, and make efficient use of network resources. It's a > closed-loop process that continuously adapts to changing traffic demands. > > Second, many modern bandwidth-engineered TE networks have paths exceeding > 15 hops. One or two loose hops might be sufficient for basic traffic > placement, but that's a different problem from bandwidth engineering. When > booking bandwidth, the actual links traversed matter. The same is true for > controller-computed paths, similarly, often there is a need for > considerably more than two segments or waypoints to satisfy bandwidth and > other TE constraints. This isn't simply a question of distributed versus > centralized path computation. > > Third, failures are where bandwidth management really matters. When a link > or node fails, the remaining network resources can become severely > constrained. Getting traffic around the failure is only half the problem. > You also need to determine whether the surviving paths have enough > bandwidth, which traffic gets priority, and what happens when there isn't > enough capacity for everyone. Without admission control and bandwidth > accounting, it's easy to turn a single failure into congestion across a > much larger part of the network. This is something I think gets overlooked > in discussions about SR protection and restoration. > > And, of course, loose-hop forwarding can also introduce transient loops > during convergence. Yet another reason why controlling the actual > forwarding path matters. > Thanks, > -- Colby > > > *From: *Andy Smith <andy@arrcus.com> > *Date: *Thursday, November 12, 2026 at 8:17 AM > *To: *Robert Raszuk <robert@raszuk.net> > *Cc: *mpls <mpls@ietf.org>; TEAS WG <teas@ietf.org>; SPRING WG < > spring@ietf.org>; Zafar Ali (zali) <zali@cisco.com>; Andrew Stone (Nokia) > <andrew.stone@nokia.com>; Vishnu Pavan Beeram <vishnupavan.ietf@gmail.com>; > Barth, Colby <jonathan.barth@hpe.com> > *Subject: *Re: [spring] Hop by Hop TE in SR networks (aka > draft-beeram-mpls-rsvp-sr-01) > > Robert, > > Thanks for writing this up. These are fair questions for the WGs to settle > before they spend time on the drafts, so here are answers from one of the > authors, me! > > In short: the simplicity of the drafts is that it keeps the property you > value in SR, that transit routers hold no per-tunnel forwarding state, and > add what SR alone obviously does not have, admission control that runs in > the network instead of in a controller. > > > Observation #1 > > The papers you cite answer a path computation question: given a topology > and a set of demands, how few segments does an optimizer need to place them > near the optimum. We don't dispute the answer. All five compute the > placement centrally, from a demand matrix or a model of one, and the > computation keeps the books for every link. > > These drafts address a different question: where the admission decision is > made, and who keeps the books. With RSVP-TE, each hop admits or refuses a > tunnel against what it has already committed on the link or set of links, > and priorities and preemption settle collisions between head ends, with no > controller in the loop. *Operators who prefer a controller and SR > Policies keep that option, in no way does this work deprecate any of that > or even propose to, nor are we fighting it.* Our drafts give the > operators who desire RSVP-TE feature sets support on SRv6: admission > control, priorities, preemption, autobandwidth and coordinated link > protection. > > There is also a difference in what gets accounted. With one or two > anchors, traffic between the anchors follows IGP ECMP, so per-link > accounting depends on a model of how ECMP splits it, and requires the > operator to trust in ECMP's abilities. Again, that may be and probably is > fine in many networks. In some it may not be, a strict End.X list puts > the tunnel on exactly the links that admitted it. Simple. > > Header depth no longer decides between the two. With NEXT-CSID, a strict > path of many hops packs into a few 128-bit containers, and a delegation hop > (Section 7 of draft-beeram-teas-rsvp-srv6) can impose part of the list, so > the head end doesn't have to carry all of it. > > > Observation #2 > > Agreed, and the drafts keep that property. A transit router allocates > nothing and installs no per-tunnel forwarding state. It records the End.X > SID that the IGP already advertises, and the packet is forwarded by > ordinary End.X processing. The head end holds one encapsulation per tunnel, > and a delegation hop, where one is used, holds a Binding SID. RSVP's > hop-by-hop work is in the control plane only: admitting the tunnel and > recording the SIDs. > > So the black hole you describe from MPLS-TE, a hop with no binding, can't > happen here, because no hop has a binding to lose. Link failures still need > protection, which is why the draft keeps RSVP-TE's coordinated link > protection, with the PLR reporting whether protection is in place. And we > are exploring ways to extend TI-LFA to have consideration of TE constraints > as an evolution here. > > We'd describe the result as MPLS-TE taken apart, or "disaggregated": the > IGP carries the topology and the SIDs, RSVP does distributed bandwidth > admission, and SRv6 does the forwarding. Each piece does one job, which > makes it simpler than MPLS-TE, where RSVP did all three. > > > Observation #3 > > Agreed that a reservation is admission control and not data-plane > enforcement. As you say, the same is true of MPLS-TE. The bookkeeping is > still worth having because it is shared: two head ends cannot both commit > the same capacity on the same link, which is what happens when each runs > CSPF on its own view with no accounting. > > On AI training, the case we have in mind is training across sites, where > one job's traffic crosses the metro WAN between data centers. Inside a > training cluster we agree: the fabric is built for its known capacity, and > RSVP has no part there. Those networks scale from switch radix and packet > spraying techniques and we propose nothing otherwise. > > Across the metro WAN, the useful question is not the peak of one burst. *It > is whether two jobs, or a job and everything else, land on the same links. * > Admission control answers that when the tunnel is set up: a job that can't > get its capacity on a path finds out at setup, from the network, instead of > later from slow collective operations. Reserving 100% for every burst would > waste the link, as you say, and nobody proposes it. The reservation is a > rate the operator chooses, autobandwidth can adjust it from what the tunnel > actually carries, and priorities and preemption decide what yields when two > jobs collide. The simplicity of this is obvious. > > On inference, your third plot shows what we'd expect between sites: the > WAN carries aggregates of many sessions, and aggregates of large > populations smooth out. That is the kind of traffic autobandwidth has > handled in POP-to-POP tunnels for years. > > On autobandwidth needing all unicast traffic in TE: the same condition > applies to an SR-TE controller. Its placement is only as good as the demand > matrix it works from, and that matrix has to include the traffic it > competes with. Operators have handled this in MPLS-TE for a long time, > either by keeping reservable bandwidth below link capacity or by carrying > all traffic in tunnels. The approaches differ in where the computation > runs, not in whether the traffic has to be accounted for. > > "Are you sure?" is the right question to ask of any new work, and the WGs > should keep asking it. We think the answer here is yes, for the reasons > above, and we're happy to continue the discussion on the list. > > cheers > > Andy > > > > - > Andy Smith > Arrcus Distinguished Engineer > andy@arrcus.com > mobile +1 267 253 5697 <(267)%20253-5697> > > > > > On Nov 12, 2026, at 06:52, Robert Raszuk <robert@raszuk.net> wrote: > > Dear Authors and WGs, > > I would like to make a few observations which IMO are important to > highlight when discussing RSVP-TE over SR-MPLS or SRv6. > > *Observation #1: * > > When original work on SR started there was a long in depth analysis done > on real Tier1 core networks how traditional MPLS-TE would map to SR-TE. > > We analyzed traffic demands and topologies. The result of the analysis was > very clear ... you could accomplish in the vast majority full traffic > management if you just anchor the TE LSPs across 1 or 2 hops instead of > creating hop by hop across say 10-15 hops TE-LSP. > > Yes the original concern was the depth of SRH (uSIDs were yet to be > invented at that time :) - but results were astonishingly clear. You insert > one or two segment endpoints into the packet header and the TE results are > pretty much identical to the original MPLS-TE. > > Just a few references: > > 1. Bhatia, Hao, Kodialam, Lakshman, “Optimized Network Traffic Engineering > using Segment Routing,” IEEE INFOCOM 2015, pp. 657–665 (DOI > 10.1109/INFOCOM.2015.7218434). > > 2. Schüller, Aschenbruck, Chimani, Horneffer, Schnitter, “Traffic > Engineering Using Segment Routing and Considering Requirements of a Carrier > IP Network,” IEEE/ACM Transactions on Networking 26(4):1851–1864, 2018. > > 3. Schüller et al., “Failure Resiliency With Only a Few Tunnels – Enabling > Segment Routing for Traffic Engineering,” IEEE/ACM ToN 29(1):262–274, 2021. > > 4. Brundiers, Schüller, Aschenbruck, “An Extended Look at Midpoint > Optimization for Segment Routing,” IEEE Open Journal of the Communications > Society, vol. 5, 2024. > > 5. Jadin, Aubry, Schaus, Bonaventure, “CG4SR: Near Optimal Traffic > Engineering for Segment Routing with Column Generation,” IEEE INFOCOM 2019. > > *Observation #2:* > > When original MPLS-TE was born there was no choice but to create a TE LSP > hop by hop. MPLS switching and label allocation were required hop by hop > otherwise you have a black hole in your path. If you are lucky you fall > down to LDP. If you were doing TE only and running on top of it > applications which required encapsulation (example L3VPNs) any such hole > (even caused by failures) resulted in bunch of upset customers. > > Here with SR the entire architecture does not require you to create a hop > by hop bindings. That's a feature not a bug. > > *Observation #3: * > > Many folks are fascinated with the notion of bandwidth reservation. I > reserve BW and it's mine ! Of course there is ZERO reservation in the data > plane so all we are talking about here (and in MPLS-TE) is just > control plane counter decrement (aka spreadsheet bookkeeping). > > But let's assume this is ok and let's zoom what does it really mean to us. > > I have recently heard that we need RSVP-SR for AI. Ok let's see how AI > traffic patterns look like in the time scale: > > <image.png> > So what are you going to use as BW value here ? 100% ? - well absolute > waist. And the above is just a plot of traffic when doing training of one > model. What if you want to push training across WAN and traffic from > training models will overlap ? 10 x 100% to allow BW for 10 models training > ? > > I don't think so as BW within training clusters is in the order of 800 > Gbps. > > Ok so others want RSVP-SR for AI inferencing across models service sites > ... Well the Inferencing traffic has this characteristic: > > <image.png> > So what BW are you going to allocate ? The above irregular spikes > (actually very similar to financial world transactions) do not allow to do > efficient BW reservations of any number. > > If you do it at peaks (say at 50% capacity of the link) you are wasting a > huge amount of link bw. If you do it at 25% you are likely to face overlap > even within the "allocated" chunks. Not a fit. > > So why do folks use TE ... > > Well let's consider typical POP to POP IP aggregate traffic characteristic > in core network: > > <image.png> > > Here we can see that the aggregates of large populations smooth out with > scale. But this shows another interesting point. > > Assume you reserve 25% of link capacity for your flow aggregates. Cool but > this is only good from 0:00 till 7:30 am. During the date the traffic grows > to 60% .. and this is why we invented autobandwith. > > So when headend notices traffic increase auto-bw is trying to either > increase the control plane reservations or trigger CSPF to reroute the > aggregate POP to POP LSPs. > > But this works well *only* in networks with 100% of (at least) unicast TE. > And as proven in observation #1 you can accomplish it with redirecting the > traffic to 1-2 anchor nodes instead of building hop by hop LSPs. > > *Conclusion: * > > The fact that one can ask AI to write code to run RSVP-TE over SR-MPLS or > SRv6 core does not make it automatically a cool thing. > > AI with how great the tool it is never complains and never asks the > question: > "Mr user - Are you sure ?" > > Kind regards, > Robert > > > [EXTERNAL] > _______________________________________________ > spring mailing list -- spring@ietf.org > To unsubscribe send an email to spring-leave@ietf.org > > >
- [mpls] Hop by Hop TE in SR networks (aka draft-be… Robert Raszuk
- [mpls] Re: [spring] Hop by Hop TE in SR networks … Andy Smith
- [mpls] Re: [spring] Hop by Hop TE in SR networks … Barth, Colby
- [mpls] Re: [spring] Hop by Hop TE in SR networks … Robert Raszuk
- [mpls] Re: [spring] Hop by Hop TE in SR networks … Barth, Colby
- [mpls] Re: [spring] Hop by Hop TE in SR networks … Robert Raszuk
- [mpls] Re: [spring] Hop by Hop TE in SR networks … Barth, Colby
- [mpls] Re: [spring] Hop by Hop TE in SR networks … Robert Raszuk
- [mpls] Re: [spring] Hop by Hop TE in SR networks … Andy Smith
- [mpls] Re: [spring] Hop by Hop TE in SR networks … Robert Raszuk
- [mpls] Re: [spring] Hop by Hop TE in SR networks … Vishnu Pavan Beeram
- [mpls] Re: [spring] Hop by Hop TE in SR networks … Robert Raszuk