Three under-appreciated facts about broadband

Fact #1: Broadband is a stochastic system

For every port in every packet network element, there is a buffer. The buffer is there so the transmission resource can be shared. The act of sharing of the transmission resource creates contention, as two things can ask for it at the same time, but only one can go first. This contention has to be distributed somehow.

The allocation of contention to flows can be seen as two processes: one process manages whether packets get in to the buffer; the other manages when they get out again.

These processes are “stochastic”. That means we have a system whose state evolves over time in a probabilistic way. Whilst its specific future state may be indeterminate, we can still reason about its limits and model its general behaviour.

So what?

For users: The resource is statistically shared, so instantaneous demand can exceed available supply. This “over-booking” means they have to accept some level of variable QoE, i.e. varying levels of “disappointment”, up to and including occasional application failure.

For operators: The quality and delivery cost of the service depends on how you choose to bias the “dice” of the stochastic system. You do this by appropriately designing the collective effect of the individual stochastic processes.

For vendors: The stochastic nature of the system means modelling supply and demand using stochastic metrics and algebras. Today this is not being done, resulting in unpredictable development costs, difficult sales cycles, and unplanned in-life performance problems.

For policy makers: Regulators need to understand the stochastic constraints of the systems they regulate, so they make effective policy that sits within what is mathematically feasible.

Fact #2: There is an effective performance envelope

Broadband networks are finite resources, so cannot offer unbounded performance. To put it another way, satisfying all demands at all times implies unbounded cost. So given cost is always eventually a limitation, what are the corresponding constraints on performance?

An empty network has completely deterministic properties. (It just sits there and does nothing!) The moment you add a load, applications interact with the network, and hence with each other. Up to a point, this process exhibits a kind of “predictable randomness” to the network’s performance.

We give a name to the level of network performance variability that any application can tolerate and still be deemed to “work”. It is called the “predictable region of operation”, or PRO. We can also have a PRO for the acceptable performance of a collection of applications, which is what a typical broadband service supports.

Beyond the PRO, you are at risk of unacceptable application performance.

So what?

For users: Different services have different PROs for any given application. Users need to select a provider based on it having a suitable collection of PROs for their application needs.

For operators: They need to have a sense of what the users are attempting to achieve, to know what the PRO target actually is, and to manage the system to stay within it.

For vendors: They need to be able to market their products based on being able to deliver a broader and more dependable PRO at a lower cost. This means moving beyond simplistic performance measures.

For policy makers: They need to think about the desired user outcome(s) are, and regulate towards the efficient delivery of those outcomes. In particular, they need to make PROs more transparent, so that buyers can make informed choices.

Fact #3: Performance is an emergent outcome

Broadband application performance outcomes are an emergent property of the system. This confuses a lot of people, because they (wrongly) assume that any specific outcome they experience was intentional.

This is rather like how the weather is an emergent outcome of uncountable interactions. The individual daily weather is unpredictable (at least where I live). If it happens to be sunny on my birthday I am merely grateful for the coincidence. When we decide where to live, what crops to plant, or where to go on holiday, we instead look at the more predictable long-run weather pattern, i.e. the climate.

Networks have a kind of “weather” and “climate”. As with the real world, we can’t control the moment-to-moment weather. However, since networks are man-made worlds, they can be engineered to have particular “performance climates”. This is currently atypical: most ISPs work on the basis of “you get whatever you get”.

The emergent nature of performance results in common error. Just because a network was ideal for a crop of Skype or Netflix in one season, you cannot assume that this was a conscious design choice by the provider. Past performance is not a good guide to future performance.

So what?

For users: Users face the risk of unplanned performance degradation unless the PRO is assured by the service provider to be fit for specific uses. Otherwise, it can shift and shrink as a normal emergent property of how broadband works.

For operators: There is an economic sustainability issue as planning processes often make unsubstantiated extrapolations of performance from past behaviour.

For vendors: They produce boxes that are just component parts of systems. By taking more control of the emergent system properties, and reducing performance risk, they can raise the value of their offer.

For policy makers: The exact nature of the stochastic processes themselves is irrelevant: only the emergent PRO(s) matter. In particular, regulators should avoid being drawn into defining performance mechanisms, since they can’t use those to control the emergent outcome. (Regulating the presence or absence of “priority” comes perilously close to defining such a mechanism.)