Comment by AnthonyMouse | Hacker News Reader

Comment by AnthonyMouse | Hacker News Reader

Parent Full thread

> The basic problem is that a linear increase in bandwidth for the customer results in an exponential increase in costs for the ISP

> N * (N - 1) =~ N^2 (assuming full duplex links, which would be generally correct for the internet)

1) N^2 is not exponential, it's polynomial.

2) Expanding capacity isn't even N^2, it's just linear.

If you double each residential customer's bandwidth, you "only" have to double the capacity of your uplinks to other networks. The number of endpoints or total nodes in the network is entirely irrelevant, because you have the same 50Mbps whether you're drawing it from one AWS server or a thousand BitTorrent peers. The only way you would get the result you're assuming is if every endpoint had a fully meshed connection to every other endpoint, i.e. each customer gets 50Mbps to each Amazon server, a separate 50Mbps to each Google server and a separate 50Mbps to every single other user on the internet, for a total of several hundred billion Mbps for every customer. That's not how it works.

And in practice it's even less expensive. It's sub-linear. Because just giving everybody a connection which is twice as fast doesn't mean everybody is going to immediately double their usage. A large fraction of users, given a faster connection, will do with it only what they currently do with the slower connection. Their pages will load faster but they won't load more pages, so the load they put on uplinks to other networks will remain the same.

There are no exponentially increasing costs. There are only the same costs as there ever were: Cisco et al come out with a new router that puts twice as many bits through the same piece of cable and if you want those bits you pay them a fixed cost and swap out your old model for the new one, and in another few years you do it again.

> Because just giving everybody a connection which is twice as fast doesn't mean everybody is going to immediately double their usage.

My understanding is that peak load is the issue, not total data transferred in a month. If Netflix downloads/buffers video content as fast as it can, and people mostly use Netflix at the same time of day, then giving them more bandwidth is a potentially big problem

Peak load is what you have to design capacity for, but that is hardly unique to Netflix. They seem to have it sorted for FiOS/U-verse/etc, don't they?

The argument they might make is that their services keep the data closer to the users, but that's such a cop out. They could do the same thing for Netflix with transparent proxies if they wanted to (but the result would be to make Netflix more usable instead of less, which is adverse to their interests as competitors). And in any event the uplink to a large peer is a totally inconsequential part of the cost of operating a network. The thing which is expensive is upgrading the last mile to handle a large amount of video traffic, but that cost is the same whether the traffic is Netflix or FiOS TV.

It sounds like what we need is the equivalent of local loop unbundling for internet services: http://en.wikipedia.org/wiki/Local-loop_unbundling

In the UK, I believe the main practice revolves around broadband, whereas in US it's all around phone (VoIP or otherwise). If you could enable competition within the same last mile service, you'd a)give a definite incentive to make sure the entire network delivers the best experience and b)have alternate sources of revenue to maintain the last mile network.

The current duopoly that exists SoCal in most areas (1 cable provider and 1 DSL provider) is a really terrible consumer experience. For those who have access to FiOS, it's a tad better, but not by much.

If you had access to 10 providers at once with comparable services, there would be real incentive for them to offer the best possible experience and lowest possible prices to consumers... but of course that would hurt corporate profits and so on and no politician getting those donations wants that...

What we really need is the equivalent of Glass-Steagall for internet services. Prohibit the company that operates the last mile from owning or being owned by anybody that offers end-to-end connectivity or over the top services to end users.

The problem with just local loop unbundling is that the last mile provider has the incentive to favor their own services. They can operate their own ISP division at a loss or with zero margins (more than made up for by profits in the last mile division) which disadvantages competitors who have to lease the last mile and thereby maintains the status quo of no competition.

Thank you. I'm glad to see that hn still knows what exponential growth is. It really bugs me when people abuse the term.

I think people get confused by the Startup rule of Descriptive Speech which states that all growth is exponential, all interfaces are gorgeous, and all teams are composed of A players.

And all design is beautiful.

Breathtaking is the 2014 beautiful.

The thing is, it's not even close to doubling. I recently exchanged my 30Mbps ADSL to a 100Mbps fiber. In practice this means a jump from 0.8Mbps (yay long copper links!) to 50Mbps, which is most probably limited by the core. Thus, for vast residential areas (this is near the center of a 80k people town in a metro are of about a million), we are talking about multiplying the last mile by a factor of roughly 100x.

Ask yourself whether that ISP allowing speeds to languish for more than a decade such that they provide one ~100x increase instead of six or seven generations of ~2x increases has made their total costs lower or higher. The fact that they've been screwing over their customers to save a buck for a long time is not a thing that should inspire sympathy for their expenditures.

What are the 6 or 7 generations of upgrades that you think should have been implemented between 30mbs ADSL and 100mbps fiber?

What 100x performance difference do you think exists between 30Mbps ADSL and 100Mbps fiber?

The answer is that there exists a box that mounts on a telephone pole. The purpose of the box is to terminate hundreds/thousands of DSL lines and connect them all to the central office with a single fiber optic cable. It allows you to get 30Mbps out of your 30Mbps DSL instead of getting .8Mbps, because the box is close enough to your house to get full speed unlike the central office. It saves the phone company from having to string a new strand of fiber to the premises of every individual customer who lives too far away from the central office for high speeds over twisted pair copper.

The first generation would have been actually installing the boxes, which they apparently never even did. Then you periodically upgrade the boxes to provide faster speeds as the technology improves and/or the cost comes down, e.g. from 1.5Mbps DSL to 3Mbps to 6Mbps to 12Mbps etc.

Giving everyone a guaranteed connection is a theoretical networking problem. The brute force solution as you've mentioned, is a crossbar switch, which is implemented in O(n^2).

However, theoretical networking problems like this have been solved many years ago. The Clos Network can provide full duplex links at significantly better big-O

http://en.wikipedia.org/wiki/Clos_network

I don't fully remember the details, but I believe you grow at approximately N*log(n). The problem is that the original Clos paper was published before big-O notation was invented... and that this was just "one other homework problem" that a professor gave to me about 5 years ago when I was in college.

So my memory is fuzzy, and the math is undocumented on the internet :-(

The internet is built on top of unreliable datagrams, which means the connectivity problem is even simpler. It satisfies the conditions of a Rearrangeably nonblocking Clos network. So the Big-O is even smaller than the above networking problem.

Clos wrote his paper specifically for phone connections. You cannot disconnect people for no reason while they're in the middle of a conversation. However, unreliable IP packets can be disconnected and rearranged, allowing you to use a cheaper form of the Clos network.

Either way, (like my professor from half a decade ago...), I'm going to leave the asymptotic complexity of "Strict Clos networks" and "Rearrangeably nonblocking CLOS networks" as an exercise up to the reader.

Mostly because I don't remember the solution... As a hint, replace the crossbar switches (in the wikipedia page) with a recursive Clos network and solve for the Recurrence relation. Use a 2x2 crossbar as the base case.

Have you thought about what such an architecture would mean bandwidth-wise on a WAN network ? I mean, dear God. This would actually require MORE links than full mesh.

In other words, core network capacity (meaning total bandwidth in all links) if you have a clos scales even worse than n^2, with n the bandwidth you deliver to users. Even worse : you need as many long-haul links as you have subscribers (because your subscribers are connected to different locations).

Is there question in your mind as to why ISPs don't do that ?

Reply on news.ycombinator.com