What can the lady who controls admission and seating in a Chinese restaurant tell us about telecoms networks?
Before Christmas I had the chance to indulge myself at a restaurant in Chinatown in London. My dear companion wanted us to have a lovely dim sum together. As they stopped serving this at 5pm, he phoned ahead (he speaks their lingo) to confirm it was OK for me to go to order, and for me to then wait for him to catch up.
There was a lady at the door who put me in seating area by the entrance to wait for a table to come free, and also for my friend to arrive. I placed the order while waiting (with less than 100% accuracy, it later turned out; my Chinese isn’t so strong). Meanwhile, other people came in. Some were seated before me, some after, depending on the group size.
Eventually I was herded to a table, where my companion soon joined me for a nice meal. It was clear that this was a restaurant that had successfully managed its tables for a fast turnover of people, with all places being full.
My telecoms eye saw the restaurant as a fixed and constrained supply resource. This was being used to serve a flow of demand (with a particular arrival pattern). The restaurant was being run at saturation. You can’t build more seats when more diners arrive, and it’s not possible to size the restaurant to the instantaneous peak demand for food. Therefore, the restaurant has to perform resource allocation choices.
The lady at the entrance performs admission control and scheduling functions. She can look over the restaurant, see how many are waiting, which tables are empty, who is looking like they are about to leave. From this she can make rational business decisions over how many people to buffer, who to send where and when, and whether to turn people away as the restaurant (including the waiting area) is full.
The central resource allocation challenge of telecoms is that everything moves at the speed of light, whereas in the physical world the speed of sound is typically the limit. Even this is aspirational in the hospitality industry! That means signals about state of the supply resource can’t move faster than the demand being serviced.
That means you can’t see the current state of the network “restaurant”, and typically you don’t really remember who was sent to “dine” and when. As a result, the choices of which people to “send to tables” and when are highly sub-optimal. We often allow far too many people into the “restaurant”, or far too few. We create dissatisfied diners, as they turn up at tables that are already fully occupied when they had been promised one.
The way we ensure “diners” are kept happy is to over-provision the resource so that tables are kept very empty. Whilst land prices and rents for telecoms networks may not rival Soho in London, they are still expensive. Because of our poor scheduling, telecoms networks use far more capital than is strictly necessary. This can often be an order of magnitude extra in access networks.
Now we have to be very careful of anthropomorphic analogies between packets and people. But the basic idea is true: the control theory used in the physical world breaks down in the virtual one. For instance, TCP is not really “congestion control”, since the contention phenomenon happens much faster than the control loop can operate; it is more a kind of “collapse control”.
It is not just the technical mechanisms that don’t readily translate from physical to virtual. The same is true of quality management systems. You cannot directly apply the ideas of lean manufacturing, but have to re-interpret them. Toyota could stop the production line when it found a quality error, as the cars were inching along. You can’t do that in a telecoms network where there are trillions of state changes per second.
Indeed, the problem in telecoms is even harder than that. We have internetworking, so it’s like admitting diners into a restaurant in Streatham, and they go out the back into a bus laid on by the restaurant that takes them to Soho where they get to eat! The lady at the door in Streatham can’t know how many people are also heading over from Stratford at the same moment.
Furthermore, we can now see how throwing ever more “speed” at the network makes the problem get harder. As we ship diners to their tables on supersonic scooters rather than bumbling busses, the state of the “diners” in the restaurant changes at an ever greater pace. This makes the resource allocation get harder, as someone coming out of the restaurant fuming about lack of tables tells you less and less about its current seating situation.
These distributed resource allocation problems mean we need new kinds of systems to perform admission control and scheduling. These systems need to understand the “diners” and construct probabilistic models of their demand. They must then carefully control how supply is used: how many to admit into the waiting area, how long to make them wait, and where to send them.
What is critically missing in networking is a better “Chinese dim sum restaurant door lady”. She will keep in mind the essential essence of who went previously and how hungry they looked. She can then make good choices based on a model of the state of the restaurant in her head, rather than having to cast a glance over her shoulder.
For the latest fresh thinking on telecommunications, please sign up for the free Geddes newsletter.