A “Packet Time” standard for the telecoms industry

The telecoms industry is automating a failed technical and business model. An external reality insists that change must happen to a common “Packet Time” standard. Your choice is to jump or get pushed (or maybe squashed). Which do you want?

A Packet Time standard for the telecoms industry

Earlier today I sniffled my way through a speech at SDN NFV World Congress on why the telecoms business needs to rethink something core to its existence: the nature of its resource and how it models it. Here is an edited version of my notes made to run a bit more clearly than my head did when on stage recovering from a cold.

There are no slides for this talk. I am not important. Please pay attention to what I say. It is critical to your welfare.

In the 1990s we saw the dotcom Internet boom, and now we have a “fourth industrial revolution” unfolding with IoT and the industrial internet.

However, the dotcom boom was nothing compared to earlier industrial revolutions. In the 1940s the Railway Mania far exceeded telecoms as a share of GDP.

That industrial revolution was only possible because of one thing: we finally agreed on how to align the timetables for the trains by the introduction of Railway Time and GMT. Before then, everyone had a chaos of different local timing standards, with noon being whenever the sun was overhead.

The common railway timetable allowed the synchronisation of the use of the underlying resource, which is the train track and the train paths that go over it. It is effectively a circuit-switched model.

Packet networks are stochastic systems, based on probabilistic processes. However, our resource models do not reflect this fact. It’s as if we’ve confused the transmission rails for the packet trains and forgotten we need a common Packet Time.

That means there is an external reality you HAVE to align to, and you MUST change your resource model for SDN if you wish to succeed.

This is not an optional matter.

Indeed, you have to change resource model for three reasons:

  • The current resource model is not scientifically sound.
  • If you continue on this path then bad things will happen, and good things will definitely not happen.
  • If you change to a scientifically sound resource model, the bad things will not happen, and the good things might happen.

What is a resource model?

Before we go there, let’s first get a basic definition straight. What is a ‘resource model’?

When we open a water tap or faucet at home, we express a demand for water. The water pipes supply water to meet that demand.

Our resource model has two elements: the volume of water (in litres, say), and whether it is fit to drink. We define the resource flow in terms of litres per second. (The quality is already set by law.)

When we cook with gas, we demand an energy supply. Our resource model might be expressed in British thermal units, or kilocalories, or megajoules. We define the resource flow in megajoules per second.

Broadband networks are systems that make performance for distributed computing applications. The network is a finite resource. It is a system that system of supply for the demand for timely computation. It offers the ability to “translocate” (i.e. copy information) between those computations.

SDN (and NFV) place the translocation (and network computation) supply under software control. This has to dynamically match supply to demand.

So what’s the right unit for performance?

Not a scientific resource model

A scientific model is one that is predictive, with understood fidelity to reality; it makes use of mathematics, not misuses it.

Today’s network resource model (used in SDN) is “quantity with a quality”. It is the “classical” bandwidth model of resource that makes quality ancillary to quantity.

However, it suffers from a basic problem: you can’t meaningfully “add” and “divide” the resource. The units don’t add up! They might do so as numerals, but it’s like adding “3 o’clock” to “7 o’clock”. It’s not a meaningful operation; you can’t just declare the answer to be “10 o’clock”. Similarly, adding “bandwidths” doesn’t conserve quality; it’s a simple silliness.

Indeed, we have multiple different units with unknown couplings: throughput, jitter, and loss being typical metrics. Averages of quantity (like litres per second) don’t capture quality (like is the water fit to drink).

But the value we deliver is quality, since performance depends on the timely arrival of enough information. If it isn’t timely, you can’t make it up on quantity! Furthermore, the cost is driven by quality. If you are willing to wait forever to get the resource, your marginal transport cost is zero.

So why are we using metrics anchored in quantity?

The unavoidable truth is that packet networking is statistical multiplexing that introduces variable quality. The customer experience is the continuous passing of instantaneous moments of quality. That means we need a probabilistic resource model to reflect this.

The bad future that lies ahead for SDN

If we continue with this bandwidth madness, several bad things will happen.

Because we are using metrics that are poor proxies for QoE, we cannot engineer predictable application performance. The metrics are disconnected from the user experience, so cannot result in an assured experience. The user is carrying an excess risk of unexpected failure, which makes the service worth far less.

Because our units don’t “add up”, we can’t model the properties of complete systems, since we don’t really know the “safety margin”. You only find out if your SDN orchestration failover works after the hurricane wipes out your data centre. The lab isn’t a good guide to the real world, where everything is stretched over thousands of miles of fragile glass and damp copper.

Even worse, because we insist on switching “bandwidth”, every change made under software control potentially induces a discontinuity in quality. We call this unpleasantness “non-stationarity”, and it makes your users’ applications fail.

This destroys value.

The scientifically sound resource model

The good news is that a better alternative exists. This new resource model is a “quantity of quality”. You can think of it as being the “quantum” model of the probabilistic resource.

In this model, quality comes first! It captures the essence of what matters: the delivered performance to the end user.

The basis for this model is is a new branch of mathematics called ∆Q. You can think of it as the “complex numbers” of probability. [Speaker inner panic: “Crap, I hope they know what complex numbers are or else I am in trouble.”]

This new mathematics extends the idea of randomness to include “observable non-events”. It unifies unifies continuous phenomena (like packet delay) with discrete phenomena (like packet loss) into a single model and metric.

∆Q also forms an algebra and calculus. That means you can reason about performance and budget it. So just like you don’t worry about the conference centre and hotel falling on your head, you don’t need to worry about unexpected failure of your SDN-driven network.

The ∆Q framework has a known and very small infidelity to reality, hence it is genuinely scientific.

A better outcome with science and engineering

When we adopt quality-centric metrics, we have a chance to engineer experiences and sell assured quality for cloud access, which drives more revenue.

We can construct networks with a known safety margin, which makes them more valuable to network operators who no longer have to carry the risk of unplanned failure due to poor engineering practises.

We can not only offer predictable performance with high customer value, but also run networks with very high resource use at the same time.

What’s not to like?

How to get from here to there?

These ideas have been applied in reality. They were developed for clients like the US Department of Defense and Boeing. They have been used to optimise getting data out of the particle supercollider at CERN. They have also been used at many tier 1 fixed and mobile operators.

My colleagues have built the world’s first “fMRI” scanner for networks, and first quality-assured broadband solution. It’s a breakthrough.

The drawback is that today, none of this in the textbooks or taught in universities. It ought to be in undergraduate computer science courses.

That said, the core mathematics is all available for free. I am ask you to help me to spread the word that the maths and science is available, and to help me to disseminate the tool chain.

A closing thought

If you are working on spectrum policy, we don’t have arguments about physics. We sorted out the mathematics (with complex analysis) in the 17th and 18th centuries, and the physics of electromagnetism (with Maxwell’s wave equations) in the 19th century.

With computing, people like Church, Turing and von Neumann put in place a theory of computability in the 1930s, before we built digital computers.

With packet networking, we have things backwards. We’ve been building them without a fundamental theory. ∆Q supplies the missing “theory of information translocation”. It’s the “quantum mechanics” of packet networking, dependent on a new “complex numbers” of probability.

We need a new industry-wide “Packet Time” standard. You have to change if you want to prosper. It’s not because Martin says, but because reality insists.

As Richard Feynman said in a 1979 New Zealand lecture on physics and quantum mechanics:

“[People say…] ‘I don’t believe it. It’s too crazy. I’m not going to accept it.’…

You’ll have to accept it. It’s the way nature works. If you want to know how nature works, we looked at it, carefully. Looking at it, that’s the way it looks.

You don’t like it? Go somewhere else, to another universe where the rules are simpler, philosophically more pleasing, more psychologically easy.

I can’t help it, okay? If I’m going to tell you honestly what the world looks like to the human beings who have struggled as hard as they can to understand it, I can only tell you what it looks like.”

Thank you.

I would like to thank Mark Lum and the crew at Layer123 for inviting me to speak. It was a genuine pleasure to attend and there were lots of really good people and interesting talks going on. I look forward to spending much more time at the event in future.

A Packet Time standard for the telecoms industry - cake

Apparently Escher was keen on high or low perspective points. Personally, I am keen on caramel latte and chocolate caramel cake. If there are typos in the above, it’s closing time at the museum and I have to get a train to Amsterdam.


For the latest fresh thinking on telecommunications, please sign up for the free Geddes newsletter.