Slashdot Mirror
Breakdown of Bandwidth Costs?
WCityMike asks: "What is the origin of the cost of bandwidth? For instance, if I'm being charged for an apple, I know that, theoretically, the cost of that apple is going towards the purchase of apple seeds, the land on which the apple trees are grown, the fertilizer and water that helps the trees grow, and the salaries of those who pick the apples, clean them, box them, and send them to market. When an Internet provider charges someone hundreds of dollars in bandwidth costs because they were Slashdotted (or Farked) and their bandwidth use shot up, what costs have the Internet provider incurred, and why does it cost them what it does? Is there usually any sort of markup going on along the line, or are people just passing along their own expenses down the line to the end user?" It would be interesting to note the most important factor contributing to bandwidth costs. How much of the total costs are tied to infrastructure versus the human component (technicians, sysadmins, technical support and so forth)?
...the problem is burst. Period.
When you buy an apple, you can eat the whole apple at once (tossing the core into the compost pile when you're through), or you can cut the apple up into slices and eat those slices over a few hours' time. If you eat the entire apple in one sitting, it doesn't cost any more than if you'd cut the apple into 16 slices and eaten one per hour.
Bandwidth isn't like an apple.
Bandwidth is purchased in one of three ways (or sometimes several or all three ways at once):
1. A connection with a maximum of X amount of throughput at any given time, "unmetered" (ha ha)
2. A connection with a maximum of X amount of throughput at any given time, plus a restriction on how much throughput can take place in a given month
3. A connection with a maximum of X amount of throughput most of the time, which can increase on the fly to some multiple of X for short periods of time, with a restriction on how much throughput can take place in a given month
The third type of connection is usually known as "burstable." What it means is that the ISP sells you a line which is physically capable of, say, 3Mbps transfer; but they sell you this line with the understanding that 99% of the time you're only going to be using a maximum of, say, 1Mbps of that capacity. You're allowed to "burst" over your 1Mbps average up to 3Mbps.
Some ISPs restrict these "bursts" to certain lengths of time, some restrict the "bursts" to certain amounts of data (e.g. you can burst to 3Mbps for 1GB at a time). Either way, you're typically given a certain allotment of "bursts" per month. Any more "bursts" and you pay extra.
When your site is Slashdotted, you generate hella "burst," hella quick, and hella strong. As a result, your hosting company may wind up exceeding the allowable "burst" (either in time or in bytes) according to their contract.
Unlike the apple - where you can "burst" it all into your stomach at once with no penalty - bandwidth is sold on the contingency that it's going to be used over time. Use it all at once and you get fucked. That's pretty much it.
For my organization, about 45% of the customer's cost goes to pay for bandwidth. The rest is mostly people costs.
- We provide very limited end-user technical support.
- We provide a very specific service. We don't try to do "anything that could bring in a buck". I think a lot of ISPs get into the trap of "anything for money"... which results in a LOT of hidden costs. This should be MBA 101 stuff, but it's amazing how trapped people get doing things for zero benefit.
- Before we implement anything, we look at the potential costs and benefits.
- Everything has to have a complete process in place before we move it into production. Otherwise, on-going support costs skyrocket.
- We have a very clear contract with our customers. We trust them; they tryust us; they don't misuse us, we don't misuse them.
This is simply basic IT business stuff. But most small bandwidth resellers (and many large providers) fail at it.
Not all ISP's peer for free with other ISP's. In a lot of cases, they have to "settle up" for their peering costs at the end of the month, per megabyte, with their peers. UUNet is big on doing this with their peers and it's where they make 50% of their income...
Cogent and AOL just had a fallout about this - Cogent felt that they should have free peering with AOL but AOL felt Cogent should pay for usage because most of the traffic going to AOL was for Cogent customers and not vice versa.
(see 'Peering' Dispute With AOL Slows Cogent Customer Access and Paid Peering )
So yes, the short answer is that upstream ISP's do pay per Mb in a good amount of instances if they aren't a huge transit hub. Obviously some ISP's aren't subject to this and just charge by the MB figuring the more traffic you're getting, the more money you're making from their connection.
Sig (appended to the end of comments you post, 120 chars)
Anyway, my bandwidth was paid for. I had a full T1, Sprint didn't care how much traffic I sent across it, they were set to cover it.
He was very sorry about it. It cost me about 2 hours of downtime for the rest of my customers. It took me about an hour to respond once I found out there was a problem and pulled the plug on his box. I issued refunds to customers that were affected.
I charged him the cost of those refunds, plus $35 for an hour of my time.
Note that this probably didn't cover all of my costs. I wasted most of a day adjusting bills of those affected, which I didn't charge for. The agreement I had with this customer was special. He was allowed to use all extra bandwidth for his peering box, but was not allowed to infringe upon the rest of the customer's requests. This was before QOS packet stuff so it was kind of a manual 'keep on eye on the MRTG graphs and make sure were aren't maxed out' kind of arrangement.
First off, an increase in load does not mean you need to increase the number of people who work for you (sysadmins, techs, etc.)
Second, costs are obviously passed to end users with a mark up (however little).
Third, cost of bandwidth DECREASES AS BANDWIDTH ITSELF INCREASES. At least with ISPs. ISPs who buy bandwidth in bulk actually reduce their cost per Kbps (that is, we get charged a lower per Kbps rate if we buy 45Mbps outright than 64Kbps). For example, we used to pay $1600 for a 64Kbps link but now we pay only $6000 for a 2 Mbps link (we're outside the U.S. so this includes physical connections) instead of $51200 ([2048Kbps / 64Kbps] x 1600).
HOWEVER, buying in excess usually result in part of the bandwidth being unused but the ISP still pays for it. That's where the numbers game starts, where the ISP divide the cost of the bandwidth plus a mark up, to get what they charge customers.
Obviously they can't charge customers on a per Kbps rate to recoup costs otherwise the customers would be up in arms. Now here's a secret (ssshhh!), most Internet traffic is bursty, that is, customers don't always utilize the entire bandwidth (at least, not at the same time) that the ISP guaranteed(?) - you access a webpage, read it (while you read, you don't utilize bandwidth) and access another page. ISPs rely on that fact when coming up with prices. That is the reason most do not like it when you upload stuff in P2P nets. Most probably they just use the RIAA, DRM, whatever, as an excuse to block access to P2P. Why? Because the minute you offer upload access over P2P nets, chances are high that you'll be utilizing whatever bandwidth they promised you. They don't like that because if a significant number do that it'll throw off their numbers (remember, they didn't expect you to actually consume the entire bandwidth so they didn't charge you based on a per Kbps rate). Noticed this trend lately with so-called broadband providers who "guarantee" you so-and-so amount of bandwidth/speed?
You asked for the breakdown of bandwidth cost. Yes, most of it is caused by having to get additional bandwidth (larger pipe to handle larger traffic) but it is actually holding on to that large bandwidth (with very few users paying for it) which is the cause of you getting slapped with a huge bill if you go over the allotted amount. Their cost would be lower if they didn't have to keep excess bandwidth to give to you when the time comes that you actually need to burst over when you're slashdotted.
We solved this problem of having to offer competitive prices while still being able to pay for bandwidth by charging on a per-MB basis (with a minimum of course). We give enough free MB to cover a month's worth of surfing/browsing (done thru extensive study of usage patterns) but the minute customers go to P2P nets, we hit them with an excess usage fee. This limits them from keeping their connections 24 hours doing nothing but downloading mp3's and mpegs.
And we don't even have to hide behind RIAA or DRM.
ALL our customers btw, are satisfied with this arrangement. Only those who go to P2P nets and download recklessly are the ones complaining (until now, all those who exceeded our allotment by a large margin have ALL been caught using Kazaa, iMesh or Morpheus).
Yes. Only the big boys have free peering between themselves. And, even they have costs associated with peering: cost of colocation at a peering point, bandwidth between the two service providers, etc.
Or do they have 'back and forth' arrangements where the ISPs only pay for the difference between in and out traffic?
Some do this. However, it has always been a hot button issue with ISPs. Some ISPs, e.g. UUNET, have a higher percentage of web servers. While others, e.g. Earthlink, have a higher percentage of users. Who exactly is providing more value to the other. Is Earthlink supplying the viewers to the web sites? Or, is UUNET supplying the web sites to the viewers?
I feel like picking a fight with everyone who thinks they are right. - Rainmakers
Thought there have been several responses relating to burstable pricing, it's important to understand that burstable models are more of a marketing concept than a network costing one.
/can/ permit overbooking to occur.
Considering that IP service is in the class of packet sharing telecom - meaning you do not have a fully reserved, "100% yours and only yours" pipe from location A to location Z (end to end), your costing models have to be a bit more complex than adding the sum of the parts. But it should probably start there. Packet share != Overbooking, although packet share networks
To explain costing on a more common tier two or smaller level (tier one costing has much more to do with transport costs and a bunch more variables not interesting here), take your typical tier two or smaller ISP who we'll say is connected only to a single upstream, and that upstream provider is a tier one (multihomed at x+ bilateral and/or multilateral facilities; e.g. Sprint or AT&T Worldnet).
That ISP will have:
- an egress transport cost: e.g. a T1 to Sprint and perhaps a local loop cost to get to the tier one carrier; sometimes this is bundled by the tier one provider
- a facility cost: router, colocation, switching and all the associated things to figure out what to do with customer traffic and send it where
- an ingress transport cost: this is the transport cost from the customer to the ISP (e.g. T1 to a local frame relay cloud from the customer, then another portion from the cloud to the ISP, or a similar model in DSL, or a fixed point-to-point line)
Plus applicable switching/routing/data center/network core/backoffice/customer support/billing costs.
Then the fun comes in calculating an aggregate the ISP wishes to use in predicting how many seats it can sell on its T1 until the plane can get full - understanding that if it oversubscribes the T1, it may have the possibility of bumping passengers (such as slow performance; people getting less than what they paid for).
(Note: You'd be surprised how many smaller ISPs don't even understand the concept of aggregation - many rural ISPs in our parts think Internet comes in T1 and put 500 DSL home connections on the T1 without thinking twice. "Order another T1? Why? We already have Internet!" Plus there are other reasons they never go past a single T1 - their Cisco 2600 only has two ports (T1 in, T1 out), they don't know how to load balance more than one T1, and fundamentally it's too expensive for them)
Aggregation becomes the process of determining how to share that cost, since it's a shared network. It includes variables like time of day (business customers will usually demand more bandwidth during business hours; residential after hours, etc.), whether the customer has a bandwidth guarantee in their service level agreement, etc.
After all of this calculation and checking of assumptions, the ISP may calculate it into a cost per Mbps as we do.
so they have to overbook. It's all about getting the most out of what resources you have.
Not always true. Overbooking can be perceived as ineffective loading - an error, in otherwords, to be avoided. Your power utility doesn't run a constantly overbooked network - that'd have serious regulatory and technical consequences. But when you've got very little backoffice expertise in a smaller ISP, they won't know how to develop traffic models and purchase effectively.
Overbooking can be avoided both by developing good subscriber models and by enforcing them (ugly reality: if you're using 1 Mbps up/down sustained and are paying $35/month, you're not paying for what you're consuming and you'll cause a problem somewhere along the line). Additionally, you're going to see better aggregation and closer matches to "what you pay is what you get" as the last mile part of the business matures. Much of the success of an ISP is in how it buys and how closely it matches what was bought with what was sold - you can't run a $2 all-you-can-eat prime rib buffet, nor can you charge $10 for one that has nothing but dinner roles.
Really, overbooking and the opposite end of the spectrum (giving excessive bandwidth without charging for it) are mistakes made early in the development of the business. Successful providers will tighten up both ends, providing the service people ordered and paid for, and smart customers will understand that in order to get what they paid for, they shouldn't also expect a free lunch.
*scoove*
Let's look at cost first. Typically, service providers think of costs as some percentage of the dollars coming in for a given service. They break these down into several categories. Although each company is different, here's a simple example:
- Cost of goods sold: These are dollars that the companies spends to provide the specific service. More on this later.
- "SG&A": Sales, General and Administrative. Otherwise known as "overhead" -- cost associated with running the company that aren't tied specifically to a service (management salaries, office space, etc...). Sometimes sales salaries and marketing are broken out from this on a per-service basis, other times they are lumped into the general company overhead costs.
- Margin: This is the money left over after the company's costs are taken care of. If the company sells everything directly at retail price, this is equal to profit. However, if a company has resellers or other types of "sales channels", they will be paid a certain percentage of the margin for selling the product. Discounts also come out of this chunk.
Very rough rule of thumb: cost of goods sold should *never* exceed 50% of "retail" (undiscounted) revenue in the telecom / service provider space. Of course, if a company is desperate for business they may give away very large portions of the margin percentage in the form of discounts, etc...The "cost of goods sold" for a typical hosted dedicated server generally break down into several major categories, that may or may not be broken out as different "services":
- Hardware cost & software licensing: This is the fixed purchase cost of the dedicated server itself, and any associated software licenses. This is typically paid directly by the customer up-front, or broken down into lease-like payments and buried in a flat monthly fee.
- Data center: These are costs of having the server sitting in the data center, including lease cost of the space, supporting rack and LAN infrastructure, electricity, cooling, security, etc... Typically these costs are figured based on averages per square ft. of server room space or by rack space.
- Operations / management: These are the costs of management services provided for the server: NOC staffing, management / monitoring system cost, tape backup costs, etc... In the case of colocation, this may be minimal or non-existant, while it's very significant for dedicated / managed services, since you're paying for system administrators to upgrade your OS, apply patches, etc...
- Customer service / support: Costs for the right to call someone on the phone and get help with your service. This primarily includes call center infrastructure & support staff salaries. Again, may vary widely depending on the level of service selected.
- Bandwidth: At a minimum, this is the cost of the routing infrastructure, the cost of the WAN part of the NOC staff and systems, and the monthly costs paid to the upstream providers for big dedicated pipes. Your average hosting company is not running their own fiber or even buying dedicated circuits: they are buying IP transport from a large ISP, and pay for a certain amount of bandwidth. Usually, they'll lay in a very large fiber pipe to the ISP's local POP, and then activate additional bandwidth as needed. I can't break down the long-haul ISP's costs for providing that bandwidth, but presumably it breaks down similarly.
Now, a few words about pricing, and specifically bandwidth pricing, since the poster was interested in that. Aside from bandwidth, the costs above are reasonably predictable from customer to customer and month to month. The more predictable a cost is, the more likely a service provider is to lump it in with a bunch of other costs and charge a flat fee. The less predictable a cost is, the more likely it's going to be metered and broken out.In general, customers demand / want flat, fixed, predictable prices. This is why ISPs charge $19.95/month for dialup rather than $.05 / hr. Of course, at some level the ISPs price for dialup is proportional to actual use, but they've figured out that, on average, they can make money at a $19.95 flat fee, and customers are willing to pay it. Of course, customers who use less are subsidizing the service of customers who use more. But as long as the *average* cost remains low enough that the low-usage customers feel they are paying a fair price, it works out.
Server bandwidth is a case where it doesn't work out. The amount of bandwidth that a single static web server is capable of consuming is quite stunning, and goes up along with Moore's law. Many hosting customers choose a dedicated server not because they need gallons of bandwidth, but because they have some sort of custom app or want full administrative control of their system. These folks aren't willing to pay a price that would cover the "average" cost of bandwidth across all an ISPs customers, which may include large streaming media systems or pr0n hosters. So the ISP meters / measures the bandwidth and charges each customer appropriately.
However, it's a bit more complicated than that. You'll notice that, if a customer buys a lot of bandwidth, they pay a lot less per GB than a small customer. The smallest customers may pay ridiculous prices for bandwidth if they exceed their plan (case in point the Farked "boobies" page that racked up $400 in bandwidth charges in two days). It's obvious that SPs are charging much more than their "costs" for the bandwidth in these cases, and you'll notice that the same thing is true for cell phone plans, etc... What's going on?
The answer is that it's all about predictability. A service provider must maintain adequate capacity for providing service. If they don't have enough upstream bandwidth, service quality for their entire user population goes down the toilet. It takes a long time to add additional capacity- new fiber needs to be run, new equipment purchased, etc... As a result, service providers are always buying new capacity in advance of demand, and to do this accurately they must be able to forecast demand. This boils down to getting accurate forecasts from their customers, which is impossible to do directly. Instead, as capitalists, they put economic incentives in place to motivate customers to predict their maximum demand accurately by pricing in tiers, or packages. Buy buying a particular bandwidth package (say, 20GB/month) you are effectively telling the provider that's how much you are planning to use, and they plan their capacity accordingly. Deviating from their master capacity plan is very, very costly for the service provider, and accordingly, the "right" to deviate from your plan as a customer is proportionally more expensive.
Another way to look at it. Customers who buy the smallest bandwidth package are the least "valuable" to the company because they are making no revenue commitments. They are also the most easily able to double, triple, or quadruple their demand from month to month because their servers are capable of consuming so much more bandwidth than they are buying. On the other hand, a large, multi-server customer is a more valuable customer, and is likely to be using a larger portion of the available capacity of their servers, and thus less likely to have wild changes in the amount of bandwidth they consume. The small customer thus pays much more per GB of bandwidth than the large customer, especially if they exceed their plan. In a lot of ways it's like the airline industry or any other industry where buying additional capacity is expensive and/or time consuming: you pay a premium for being able to use that capacity on short-notice.
Well, that's more than you probably ever wanted to know. Hope it's been educational.
-R
I run a small wireless ISP that covers a five county area, three of them in a 700k population urban area, two rural counties that border the city.
:-)
A single T1 from a reputeable provider (Sprint, UUNet, AT&T on good days) still costs around $1,000 in this part of the country. You can get slightly cheaper T1s from [IC]LECs ($800 or so) but if you later multihome with a tier 1 provider you'll end up lopsided. If you grow enough to need a significant fraction of a DS3 your cost per T1 equivalent can drop from the $1,000 mark to about $500.
I've posted this little equation a number of times before and I don't think its worn out yet.
one 256k full time music/movie trader == 4 channels on a T1 == (4 channels/24 total channels) * $1,000 ~= $160/mo cost to an ISP, dropping to $80/mo if they're a big ISP.
In addition to circuit costs I pay antenna site leases for building roofs, towers, and the like. I've got a small office and the debt on the equipment used in the network.
Unlike various venture capital created wireless beasts flopping around in my market, my company is 'organic' - it has financing (money used to buy equipment) but no funding (money used for operations). My partners and I all have other sources of income (consulting, equipment sales, very tolerant spouses, etc) and we won't draw paychecks until we've got the recurring revenue to support that.
We face one other wireless player in our space with a first generation bridged only network and a not so good reputation (I love you guys! Don't change a thing!). We also face a large, agressive cable company turned CLEC, DSL is widely available from the ILEC and another CLEC, and the ILEC has a fairly useable, cheap ISDN offering.
At the entry level for business we get between $75 and $130 for a connection up to 768k. We charge $250 per megabit with the assumption that anyone using that sort of speed is going to have VPN apps on the connection and they'll require additional hand holding.
In addition to the business customers we have one of the rural counties that is actively producing residential customers @ $29.95 monthly. I wouldn't miss them if they all dropped dead tomorrow, since they're mostly those that wore out their welcome at their local DSL provider and now they're Kazaaing up my network.
T1s are full duplex. Did you know this? DSL/cable users are amazed when they learn this fact. I like to explain it to them on a phone call rather than in person, so I can interrupt them while they're telling me a T1 can't be 1.5 mbit symetric. Think about it
So, I've got N T1s x 1.5 mbits of inflow and the same amount of outflow. The business customers use bandwidth from about 0700 - 1700, the residential customers from about 1500 when kids get home from school till about 2400 - basically two completely separate markets I can serve with the same inflow bandwidth. The ratio of inflow to outflow is about 5:1 with 'normal' customers.
The Kazaa thieves upset this 5:1 rule of thumb greatly and should really be counted as a wholesale symetric customer and charged accordingly. The CEO won't approve this charge plan, so I'm approaching it BOFH style, and anyone running Kazaa or similar services has just volunteered to be a victim^h^h^h^h^h^h subject in my various QoS experiments.
How did we decide how much to charge?
You need to understand that while I have an in depth understanding of what people are doing with the IP network I manage I *don't care* what they're doing, so long as they're not annoying the others trying to use it, or doing stuff that will get various three letter government agencies serving subpoenas to my office.
I happen to like playing with IP networks which makes being CTO a pretty fun job, but we could just as well have a warehouse full of various widgets that we buy for $2 and sell for $3 - its just an investment.
All networks built with more equipment than you can carry in one large RubberMaid(tm) storage container are investments. They depreciate a certain amount monthly, they have various expenses involved in operation including the next tier ISP cost, and hopefully, if you're doing things right, you're ending up with more $$$ on the 31st of December than you had January 1 of the same year.
So, there you have it. Its a *business*. Take a business class if this is all fuzzy to you, or better yet just start an ISP on your own and we'll see how long you think sharing MP3s and movies via the internet is a 'right'.
I am very easy to get along with, but I don't have time to waste being nice to people who are being stupid. -Theo