Rethinking Network Switches
The rate of technological change today can seem overwhelming but if you are old enough, you start to realize its pretty much more of the same. My biggest fear with tech change is that I don't want it to happen TO me. I may not be in a position to be in front of it but often you can prepare for it. True, it may whack you upside the head anyway. Welcome to the real world.
Here be dragons
I never wanted to be the one who said:
- This twisted pair (Cat3) stuff is a joke and it will never catch on.
- This voice over IP stuff will never work. TDM or death (or at least retirement).
- Why do we need an operations team?
- We need this physical server for our applications. Our applications won't work on a virtual server. (and it needs to be under my desk for easy access)
- Why would I learn a programming language to interact with my switch?
- Why would I run this script to do something that I can do in seconds via ssh?
...all quotes people have actually said to me over my career.
I think all of us have seen and are seeing revolution in "the thing" itself and revolution in how we consume "the thing".
Who is the road for?
My grandfather saw the transition from horse and buggy to automobile. That topic is fascinating in and of itself but equally fascinating is the story of how streets were redefined.
I've been reading Peter D. Norton's Fighting Traffic where he documents the change from a street being a public space, almost park-like with children playing, vendors selling their offerings, and the purview of pedestrians to it being an automotive first pathway. Traffic management, traffic cops, speeding tickets were things of the future at the turn of the century.
I can take books, movies, and music with me to the middle of the Pacific Ocean on one small handheld device?
These types of transitions have always fascinated me. Looking back I'm stunned to realize I've been a part of so many transitions already not to mention the ones going on right now.
Computers
- Mainframes to minis to desktops to laptops to my phone to my watch
- Physical to Virtual
Film/Photography
- I miss the days when I would allocate so many rolls of film for a dive trip and then continually be on the lookout for film I could buy at my destination. With abundance of resources comes cavalier waste and less thoughtfulness about each image. Abundance cheapens things. Images with film to digital, words with AI. You get the picture!
Books, Music, Movies
- One of my biggest challenges on trips was what books to bring and how much would they weigh.
Automobiles
- While we don't have the EV yet, we have the charger ready to go.
Power
- The panels on our roof can fully power our house during the day (as long as my network lab is off) and on a full charge through the night. (I live in the fog capital of California)
There are many more examples, big and small.
What has changed in networking?
As I look at all of this change it is really amazing how little networking has changed. That might sound controversial but I believe there is a core of irrefutable truth there.
Even network automation is just a new way to design, implement, and operate the network devices.
I remember installing the first Kalpana switch (an actual switch not a hub) at the NASA Program Office at JPL. That must have been around 1994 because I remember learning that Cisco had purchased Kalpana while we were doing the installation.
We have new protocols, new interfaces, we may even lose the SFP/xSFP for power efficiency, we've possibly stamped out telnet, we have cloud, we have better ways to deal with routing and East/West traffic flows, controllers have come and gone and come again.
We have better power usage, port density, faster speeds, arguably better interfaces but nothing really revolutionary. Even the Dijkstra routing algorithm remains intact but "may not be optional in some cases".
The final irony is that even though these really are incremental changes for the most part, it is still difficult to change out infrastructure. The inertia is substantial. Looking through this lens of change, I see how "maintaining the investment" has nearly frozen networking (structured cabling is complicit here as well).
I know physical limitations are involved but it is an interesting realization for me.
I feel like we have Stockholm syndrome and we've come to accept our captors without struggle, the network vendors.
What does the switch I would fight for look like?
Is there a switch I would fight for today? Not really, there are only preferences and comfort, and this is a level of apathy I am not comfortable with. So if there isn't one today, what does the switch I would fight for look like?
Ordering
No more than 3-5 clear and distinct SKUs per (a few) categories. Your product isn't better because it takes a Stanford degree to order it. There is no value in being part of the secret cabal that understands the licensing for it. I submit to you that if you have to have courses on ordering your equipment you've missed the mark.
Physical Size
I'll live with the RU standard. Let's look at a 1RU device. It should include a place to label on front and back (not top and bottom where stacking covers it up). More on this in a moment.
As much density as possible as long as I can get my fat little fingers in there to disconnect a cable. While we are talking about my fat little fingers why can't someone design a system where it is easy to add/remove connections and a connector without that stupid clip that invariably falls off which eventually unseats the cable and causes an outage.
Management
In the past, I would say a management port is a must but are we missing out on different ways to do out of band?
Recognizing the Stockholm effect, today I say an independent management path/access via different methods is a must. That is the real requirement.
Maybe we could learn some lessons from IoT? What about wireless and private 5G? There are many innovations in wireless and just like LAN they are constrained (by spectrum) but still we are seeing some impressive strategies and innovations in wireless. In campus environments you rarely see out of band management and less so terminal servers usually because of the geographic sprawl. Why is that still the case? In a world where connectivity is going to be even more important, how can we let this stand?
No hardware should come without a meaningful and resource efficient virtual version of itself for testing and validation. None of this run this for the layer 2 features and this other thing for the layer 3 features nonsense.
Front Panel Real Estate
Let's get back to room for a label. Why can't we have a small LCD display (The APC PDU I purchased 10 years ago had that) that shows the hostname and if there is any non-nominal state or information? Maybe I can add the RU and Rack it lives in. This can be important information and best added while standing in front of the rack. Let's rethink who all those LEDs are for? No one is sitting in the cage or MDF waiting to see if an amber light appears. I'd say the most valuable thing on the front panel is the switch ID so someone can plug something in or identify the switch for maintenance/replacement. Other than being able to ID a specific port (and the relevant switch in the stack or module) I would rethink everything else on the front panel other than reset/factory defaults capability and that can be done via the LCD as well with authentication and authorization.
The days of switch design for predominantly human interaction are over.
Let's think about when human interaction is likely:
-
Replace
-
Connection maintenance
-
Troubleshooting (connectivity, POE, and more)
How could we make that easier???
Testing and Validation
I want to be able to test from a client/application perspective from the switch itself in the data plane. Why don't we have this by default? The days of a network operator tossing in a switch and hoping it all works have long since passed. We forget that the installation and configuration of the switch is not the end, it is the beginning of the end user experience. It should be frictionless to do some basic testing.
- Can I get a DHCP lease,
- Can I get it from each DHCP server configured upstream?,
- What other lease parameters do I get?
- Did they forget to send me a gateway with my DHCP lease?
- Can I curl a critical web page?
- Can I save an application flow and "play it" for validation?
- Can I log in and show authentication is working?
Why is it not effortless to test this from the switch itself before it is released for end user use?
Why can't I tag my ports? I may not care about an end user port losing link but I do care if it is an uplink. I also care if it is connected to a camera, or providing power, or connected to an Access Point, and on and on.
Why don't I have rich flow data on the switch that I can use independently of a server?
Why isn't it effortless to execute and analyze a packet capture?
Why can't I test a topology and switch interaction from my switch itself?
Structured Data
We will never get vendors to provide uniform structured data. Let's just admit that and stop wasting time. With that as a given the switch should be able to normalize its data in a way I specify. If I don't care and I'm all in on one vendor OK then, take the default data structures. If you need normalization because you are multi vendor or have specific workflows then the switch should let me set my own schema, do the translation and offer it up via a standard programmatic interface.
Programmatic Access
REST, GraphQL, MCP interfaces and telemetry are a must. Can we put SNMP to bed? Hasn't it given its all?
Zero Touch Provisioning
Autonomous Zero Touch Provisioning because what does it say about your product that you need third party tools to provision your equipment consistently and reliably? This is not difficult. Cisco ACI has been doing it for over a decade. But wait....why do we need zero touch provisioning?
Intent
Why is my switch not intent aware?
What if my switch had a tuned LLM (RAG or otherwise) for the way I want my network to behave or even just enough to get it on the network so I or something else can finish the provisioning. In that world, ZTP is just part of the provisioning workflow and not a workflow in and of itself.
I could go on and on....
Let's leave this section with the following thought. Switches have pages of features. How many do you actually use?
Documentation
Let's face it, the days of learning syntax are behind us (yes think about this...we will talk about this later). My switch needs to have documentation that is AI consumable. Forget the documentation, it should come with skills and agents ready to be consumed by its own LLM or my LLM or my LLM service.
You know the argument has always been that we're never going to be able to mandate homogeneous interfaces, output, etc., because the vendors need to be allowed to do their own thing so that they can innovate. Well I'm trying to figure out where that innovation is. In the twenty years I've been involved in networking, not much has changed. I've come to realize that argument is just flat-out wrong.
What about features for each persona?
Borrowing a term now commonly used in AI, hardware design should have personas and features specifically designed to address their needs.
- I am a user of the network service provided by the network equipment. I just want it to work.
- I am a technical procurement specialist who needs to order new equipment for the campus.
- I am a technician tasked with installing, connecting, and powering, network equipment.
- I am a network engineer tasked with designing and implementing the network service my users need to do their jobs.
- I am a network operator tasked with operating the network service, ensuring optimal performance and resolving any deviations.
What features help each of those personas?
This is hardly revolutionary!
I think these asks are pretty modest. It is possible some are just physically impossible (until they are not) but my goal here is to ask why, to question the premise?
This list is by no means comprehensive. These are just some obvious ones. I'm no great thinker or futurist but I have to ask the question.
The question: Why don't we have these things?
The answer: We have the things the vendors decide to give us.
That's not a great place to be.
I realize that is not that unusual.
In 1903 if you purchased a Model A Ford, it only came in red.
I never realized that today, our switches still only come in red.