[texthompson]

If you didn't know, that 80% number is probably the result of Little's Law. That's the result where if your demand is generated by a Poisson process, and your service has a queue, 80% utilization of the service is where the probability of an infinite queue starts to get really high. People

Here's a nice blog post about the subject:

https://www.johndcook.com/blog/2009/01/30/server-utilization...

[lxgr]

This law does not apply to queueing as encountered in routers. It assumes unbounded queues and a poisson arrival process (i.e. a memoryless channel); both assumptions don't hold for packet routers and senders using congestion control (TCP or otherwise).

There is, however, a high chance of encountering buffer bloat if countermeasures are not taken at the chokepoint: https://en.wikipedia.org/wiki/Bufferbloat

Modern cable modems, for example, are required to implement such countermeasures. My ISP is at over 90% capacity and round trip times are still mostly reasonable. (Bandwidth is atrocious, of course.)

[drewbailey]

How do you monitor this? The 90% over capacity, would like to see where mine is at

[lxgr]

There might be a way using a cable TV receiver (see my other comment on this thread), but in my case, a sales rep of my ISP just told me on the phone.

[mikepurvis]

I have an older modem (DCM476) and it definitely doesn't have this or doesn't have it enabled. I have to use/tune queue management myself on the router side.

[lxgr]

Yes, it's mandatory only as of DOCSIS 3.1, and yours seems to be 3.0. (Supposedly it has been "backported" to 3.0, but that obviously would not apply to existing devices certified before that amendment to the spec.)

[eru]

To add:

If you have more control over or knowledge of your load, you can safely go higher than 80%.

Eg when I was working at Google we carefully tagged our RPC calls by how 'sheddable' they were. More sheddable load gets dropped first. Or, from the opposite perspective: when important load is safely under 100%, which it is almost all the time in a well-designed system, we can also handle more optional, more sheddable load.

As a further aside, parts of the financial system work on similar principles:

If you have a flow of income over time, like from a portfolio of offices you are renting out, you can take the first 80% of dollars that come in on average every month and sell that very steady stream of income off for a high price.

The rest of income is much choppier. Sometimes you fail to rent everything. Sometimes occupants fall behind on rent. Sometimes a building burns down.

So you sell the rest off as cheaper equity. It's more risky, but also has more upside potential.

The more stable and diversified your business, the bigger proportion you can sell off as expensive fixed income.

[stjohnswarts]

I've noticed that above 70-80% it gets pretty hard to insure that interrupt timing can be met and balanced with low priority main looping in a lot of my bare metal embedded projects.