Scalability Aspects of SR-MPLS

Henk Smit left a wonderful comment discussing various scalability aspects of SR-MPLS. Let’s go through the points he made:

When you have a thousand routers in your networks, you can put all of them in one (IS-IS) area. Maybe with 2k routers as well. But when you have several thousand routers, you want to use areas, if only to limit the blast-radius.

Absolutely agree, and as RFC 3439 explained in more eloquent terms than I ever could:

In particular, the largest networks exhibit, both in theory and in practice, architecture, design, and engineering non-linearities which are not exhibited at smaller scale.

Unfortunately, vendors always focus on creating solutions for their largest customers, and then try to shove them down the market with glitzy PowerPoints to leverage the investment. Bonus points if those solutions require new hardware1. Well, when I need to trim the trees in my backyard, I don’t care much about that beautiful robotic chainsaw used by people who cut down acres of forest per day.

Anyway, although most of us could probably live another day with a soft limit of about a thousand routers in a network, let’s throw some ignorant ideas around. I’m sure Jeff Tantsura will quickly post a comment saying “and that is addressed in this particular IETF draft,” and someone else will tell me what I’ve missed… and that’s why I like blogging so much ;)

But there are several technologies where you need to know the SID of the router at the endpoint where you want to send traffic. That means in SR-MPLS you need to advertise the /32 (or /128) of every router in your network. Plus their SID. If you do flex-algo, multiple SIDs even.

There are two aspects to this challenge:

  • You need to know the IP-to-SID mappings (example: BGP-free core or VPN services). You can solve this in a variety of ways (more below)
  • You need to know the network topology to calculate alternate paths (example: TI-LFA).

If we need alternate path calculation, we could either:

  • Limit ourselves to a single area, in which case our discussion is moot
  • Use manual configuration similar to Multi-area MPLS Traffic Engineering (a delight to manage and an awesome source of job security ;)
  • Use an SDN controller to collect topology from multiple areas, calculate alternate paths, and install them into your boxes. It might cost an arm and a leg, require the latest software update2, and take forever to fine-tune, but it might be a huge resume booster ;)

Anyway, back to the more interesting we just need IP-to-SID mapping use case. That’s no different from the ancient we can’t summarize loopbacks if we need to build a label-switched path with LSP challenge. Either you can fit all the information into the inter-area (or inter-level) part of your topology database or you use a protocol that’s been used forever to distribute mappings between whatever and next-hop addresses (hint: BGP). Obviously there’s a way to use BGP-LU to distribute SR SIDs (see RFC 8669 for details).

However, at least IS-IS has an interesting problem: the amount of information any router could generate (including inter-level information leakage) is limited3. Still not sure whether that’s a good or a bad thing :)

That starts to add up to a lot of info in your LSPDB. Your L1L2 routers might need to advertise several thousands of prefixes down. This is a practical concern. You can do things like increase lsp-mtu. Or implement RFC5311. I don’t find that an elegant solution.

Well, you want to have simple, cheap, efficient, elegant, fast, stable and infinitely scalable? You could have two (maybe three) out of those ;)

The best thing would be to have summarizable SIDs. SR-MPLS doesn’t have that. SRv6 does. So there SRv6 has an advantage, imho. But even when you have summarizable SIDs, you still need to know the mapping between /32 (or /128) and the SID. I don’t think there’s a good way to do that yet, besides advertising each individual /32 and SID combination.

Well, if you want to stay within a single routing protocol, then summarization might be the only option. If you’re OK with whatever solution scales, then you can avoid summarization until you hit the next limit: the LFIB size on your core routers (which have to know how to get to all edge routers).

You could get around that with some sort of hierarchical MPLS or (yet again) with an SDN controller, or you could give up and go for SRv6.

Do keep in mind that regardless of what evangelists would like you to believe, SRv6 is really a tunneling protocol like VXLAN, GENEVE, or LISP, and all tunneling protocols that are OK with sending their payload toward a summarized underlay prefix suffer the path liveliness problem4.

Back to Henk…

I’ve been told a few times that “routing is a solved problem”. I don’t agree. Routing is maybe a solved problem for networks with <= 1000 routers. For larger networks, you still need to use a bunch of trick and kludges to make it work. We could use a better solution.

Anyone claiming that routing is a solved problem has as much in-depth experience with routing as people claiming that quantum physics is a solved problem have with calculating the expected results of Feynman diagrams, but ignorance has never stopped thought leaders from having opinions. Then there are people like SD-WAN marketers who consider routing a solved problem because they “solved” the problem by offloading it to someone else.

However, I don’t think it’s realistic to expect a single unified solution to a complex challenge5. Solving a complex-enough task usually requires multiple tools and the knowledge and experience to select the best tool for the next step in the process. I do a bit of woodworking when I get tired of networking, and I have a garage full of tools. Obviously I could replace most of them with either a brutally expensive CNC machine, or with a Swiss Army knife. Neither of them would be optimal for my needs ;)

Selected Feedback

As is often the case, a LinkedIn pointer to this blog post resulted in numerous thoughtful comments, starting with Jeff Tantsura’s take on scalability:

There’s a number of dimensions to overall scale:

  • total size/amount of information needed to be maintained
  • ability to distribute the information efficiently and timely
  • amount of time taken to process the information at each node

Each dimension has a set of different limitations and requires different approaches to scale:

  • Clean it up! - IGP is not a dumpster truck (we have got BGP for that :)) these OSPF routes you have redistributed 12 years ago really shouldn’t be there
  • Each implementation exposes a different set of knobs for optimization/fine-tuning, don’t be fooled by common names, they behave differently, in a multi-vendor environment wrong combination could be rather painful - engage your vendor (if you use opensource - make sure you have got people who have a clue (not just python kids)), understand what’s exactly happening, test/test/test
  • There are significant implementation differences between vendors, without going into semantics of internals. It doesn’t really matter how fast is your fastest router, it does matter however, how slow is your slowest.

It also seems like “thousand routers in an IS-IS area” is an outdated number. Here’s Jeff’s take on the subject:

In a well tunned IS-IS flat L2 (dual-stacked + SR) you could go long way with about upto 3k routers, keeping SPF times low ane topology stable (long links are an interesting challenge though). When you start pushing above 5k, things might get nasty :)

As for IETF, there’s draft-ietf-lsr-isis-flood-reflection, I’d let Dr. Tony Przygienda to talk about.

No surprise, Tony quickly provided more in-depth information:

Number are somewhat higher than what Jeff mentions, at least in world’s most scalable implementations but yes, those limitations are being hit ;-)

Flood reflection stretches it to arbitrary number really without L1/L2 topology restrictions while supporting TE & SR (we may publish some drafts next on that).

SR in itself is one of the drivers for the ISIS scale explosion given it works best on something “flat” since SIDs are flat [until they’re summarized or are basically reinventing themselves as IP addresses as predicted by lots of old hands when original idea was shown ;-)] and call for lots of “flat IGP” network design. FR in ISIS itself is on the cusp of RFC, built and deployed ;-)

Tony also had a few thoughts on the implementation quality:

I can only say “just buy best quality/scale stuff you can get you hands on rather than looking for free lunches” ;-) High scale routing is hard though it seems simple and leans itself to well-meant promises based on assumptions, especially to people who read the RFCs but never implemented and more importantly fought large scale operational deployment realities ;-)

As another dimension, a network should be preferably designed in a disciplined way to contain the abstractions and summarize. Having e’thing being a SID e’one in the network needs to know is not particularly disciplined ;-)

If we could have e’thing e’where all’e’time, L2 bridging the planet would have worked just fine ;-) or even hard crosswiring it with one gigantic infinitely fast cross-connect matrix ;-) As further observation, cost of scaled-up, distributed, synchronized state in any networking technology is a non-linear function ;-)

Revision History

Added feedback from LinkedIn comments

  1. Mentioning new hardware in a blog post talking about segment routing is pure coincidence. ↩︎

  2. That would break tons of unrelated things because you can’t upgrade just one package in a network operating system. ↩︎

  3. Score one for OSPF where a router can redistribute the whole BGP routing table as a million type-5 LSAs. Hint: don’t do that at home, it might cause a short-term outage↩︎

  4. I wonder why this interesting draft never became endorsed by the working group or turned into an RFC. Must be a pure coincidence. ↩︎

  5. It looks like most physicists would agree (at least in private) with the exception of those still searching the for the Great Unified Theory of Everything. And yeah, I’ve spent way too much time reading Sabine Hossenfelder’s juicy blog posts lately. ↩︎


  1. For good old LDP there are (at least) two RFCs regarding scalability of transport LSPs:

    1. RFC 5283 "LDP Extension for Inter-Area Label Switched Paths (LSPs)" <>

    2. RFC 7032 "LDP Downstream-on-Demand in Seamless MPLS" <>

    The (expired) I-D "Seamless MPLS Architecture" <> describes an idea how to use the above.

    Of course this does not mean that your existing routers provide those features. ;-)

    1. Thanks a million. I love the first one -- we're worried about IGP and routing table scalability and thus we summarize IP prefixes, but there's no problem with LDP and we have infinite LFIB size ;)

      As for the second one, it seems to be one of those "let's push all the complexity into the network core (or at least aggregation)" things.

      Oh well, there's nothing new under the sun...

    2. RFC 5283 allows LDP to use a label for an IP FEC even though the RIB does not contain an exact match for this IP FEC. This is still MPLS, and each FEC in use requires a distinct label. This allows to reduce the routing table and thus FIB size via aggregation, but does not by itself reduce the LFIB size.

      This addresses the "we can’t summarize loopbacks if we need to build a label-switched path with [LDP] challenge."

      The use of Downstream-on-Demand label allocation allows to reduce the LFIB size by requesting labels only for required FECs, e.g., only for those PEs participating in locally configured L3VPNs (i.e., for the learned BGP next hops).

      In the context of "Seamless MPLS" this is intended to allow the use of smaller access network devices and still extend MPLS transport to the network edge.

      This does not change MPLS transport. MPLS forwarding is still based on exact match, not longest prefix match. An LSR does not split one LSP into two.

  2. > Routing is maybe a solved problem for networks with <= 1000 routers

    And as this is by far the most common case, I'm perfectly happy with the simple, predictable solutions we have!


    > Unfortunately, vendors always focus on creating solutions for their largest customers

    and that makes everything so much worse for all of us who are not that large :(

Add comment