Hello,

Anyone in here prefer 95th percentile over Average Usage.

Is Average Usage the same as "Bandwidth Valved"?

I did some search on GOOGLE on 95th percentile, and the sites that I looked at favor 95th percentile.

I did some search on WHT, and those posts that I read, tend to dislike 95th percentile.

Can you list the advantage and disadvantage of both ways of billing method: 95th percentile & Average Use ( or Valved Usage)?

- Asia

I have some information on my site that explains the differences in the way bandwidth can be measured. A lot of people don't understand the 95th percentile method and think it somehow unfair, hence the objections to it.

I would guess that bandwidth valved probably refers to a capped line

there are a lot of misconceptions about the 95th percentile.

i personally think that clients should be worried if their host _doesnt_ bill on the 95th percentile. since it is extremely difficult to plan adequate network headroom if you're billing on average.

raj

On average, how much additional bandwidth would you be billed for with the 95th percentile method? Double? Triple?

aling,

The fact that you are asking that indicates that you should be reading instead. You are suffering from a common misconception and the answer to your question is "none of the above".

Thanks for the FAQ on your site Duster, good info and very concise.

Best Regards,

Interesting FAQ. I have a couple counterpoints as a customer and wonder if you'd like to respond. These are why I'd personally avoid 95th percentile:

1) unpredictable cost - how do you plan when your hosting costs $100 one month and $500 the next

2) I get paid on transfer not bandwidth so I want to pay for transfer and not bandwidth - as an advertising based site, I get paid on the number of pages I display. Let's say I get $1 for every 1000 pages I show (I wish). If I show about 1.5 pages an hour over the course of the month for 1000 pages I get the same $1 as if I show those 1000 pages in 1 second. With 95th percentile billing I'd pay more if I showed them all at once even though the amount of data I transfer is the same.

3) can't control costs - If some script kiddie gets a crappy offline reader stuck in a loop (or intentionally targets a site) they can drive up the costs and the website owner can't stop them. Of course they can do this with a transfer limit too but it's much easier to create a big spike for a few minutes than to transfer 100G.

And ultimately given the choice of serving pages slowly during a spike and paying for being able to server them quickly, I'd rather serve them slowly. Sure others will chose to pay but if you're a website owner with an advertising supported content site you need to control costs.

My opinion is that 95th percentile only makes sense for small web hosts. If you have enough clients the traffic should average out across them all. If they all spike at once it there's something wrong.

Hello Jnestor,

95th percentile based on 5 min interval, don't you get like 36 hours "free"?

- Asia

Originally posted by jnestor
Interesting FAQ. I have a couple counterpoints as a customer and wonder if you'd like to respond.
Since you asked, sure, though what you've stated aren;t counterpoints, but declarations of your preferences with an explanation of why you've made the choice you have.

And ultimately given the choice of serving pages slowly during a spike and paying for being able to server them quickly, I'd rather serve them slowly. Sure others will chose to pay but if you're a website owner with an advertising supported content site you need to control costs.
People make choices according to their priorities. You've stated yours.

My opinion is that 95th percentile only makes sense for small web hosts.
I disagree. It has little or nothing to do with the size of the hosting company. It has to do with whether they value speed at all times over contained and predictable cost.

An important factor often overlooked is that all the above is irrelevant if the usage remains within the alloted amount.

I agree with you completely.

95th percentile makes sense for those who value a stable and speedy network.

It is a fair way of paying for usage based on capacity.

Cheers,

Raj Dutt
Voxel dot Net, Inc.





I disagree. It has little or nothing to do with the size of the hosting company. It has to do with whether they value speed at all times over contained and predictable cost.

An important factor often overlooked is that all the above is irrelevant if the usage remains within the alloted amount.

In general I guess we're in agreement. If you value speed over cost go for 95th percentile burstable. If you need to contain your costs go for capped.

Of course, many of the hosts out there are selling based on the transfer. It's burstable and they simply count the bytes and charge based on that. In this case I get the advantages of a burstable line and a greater degree of predictability as to the cost. Also as I said above, since I get paid based on page views, the more pages I serve the more I spend but I also make more.

Originally posted by Duster

I disagree. It has little or nothing to do with the size of the hosting company. It has to do with whether they value speed at all times over contained and predictable cost.


I only bring up size of the hosting company to explain how someone can sell burstable bandwidth based on the transfer amount. It's very likely that their upstream isn't selling to them on that pricing scheme. What they're doing is getting a fixed bandwidth at a fixed price. If any individual site bursts higher than normal for a short amount of time it won't really slow down the network overall.

If a host has 10 clients and gets capped or 95th percentile from their upstream they most likely need to charge the same way. If a host has 10,000 clients they can play the probabilities and charge based on transfer.

Originally posted by jnestor
I only bring up size of the hosting company to explain how someone can sell burstable bandwidth based on the transfer amount. It's very likely that their upstream isn't selling to them on that pricing scheme. What they're doing is getting a fixed bandwidth at a fixed price. If any individual site bursts higher than normal for a short amount of time it won't really slow down the network overall.

If a host has 10 clients and gets capped or 95th percentile from their upstream they most likely need to charge the same way. If a host has 10,000 clients they can play the probabilities and charge based on transfer.
I disagree. You are overlooking some very important facts. Many (probably most) hosts play the law of averages and oversell transfer allowances knowing full well the great majority of their customers will not approach, much less surpass, their alltoment. That same principle can be applied to bursting.

Because of that principle, no assumptions can be made on how a hosting company gets charged for their bandwidth usage.

Overselling transfer is of no consequence unless the host pays be transfer.

My point is that if you get enough sites you the peaks tend to level out much more.

I'm sure that somewhere someone has done some math to figure out how much bandwidth you need for a given number of web hosts but I'm not aware of it so I'll give an example from the Telco world were the problem is well understood.

Let's say you have an office full of people (it doesn't matter how many). You count up your telephone usage and find that during the busiest hour of the day all your people combined use 3 hours of total phone time. How many phone lines do you need to service that office? It's not 3 since you won't get the 100% utilization. At one point in the hour there will be more people on the phone. Assuming you can accept 1 call out of 100 getting a busy signal you'll need 8 lines.

Now let's say times are good and you grow your office. You study the traffic again and find you're using 10 times as much phone time or 30 hours combined in your busy hour. Now how many phone lines do you need? You might guess 10 times as many or 80 lines. Guess what - the answer is actually 42 (honest - this isn't a Douglas Adams tribute).

This effect also works for web hosts. If you have 10 clients and an average usage of 10 Mb/s during your busy hour you might need 40 Mb/s to handle peaks. If you have 1000 clients and an average usage of 1 Gb/s you might need only 1.5Gb/s to handle the peaks. These numbers are made up since I don't know the math for this but you get the idea.

Originally posted by nopzor
there are a lot of misconceptions about the 95th percentile.

i personally think that clients should be worried if their host _doesnt_ bill on the 95th percentile. since it is extremely difficult to plan adequate network headroom if you're billing on average.

raj


That is completely not true. Through our arrangement at DI we but per mbp not per gig and we get our chart set at the 50th percentile not the 95th. So we can pass on this 50th percentile measurement and easily buy more mbp and know whow much we have to buy just as someon on the 95th would.

Originally posted by aling
Average increase in bandwidth charge
On average, how much additional bandwidth would you be billed for with the 95th percentile
method? Double? Triple?

I posted this a few months back too. The 95th measurement comes out almost exactly 3x the same traffic measured at the 50th percentile. Duster, if you are still using DI and had your charts switched to the 50th then you can help me confirm this. One of our old servers pushed 330-360gig a month on the 95th, when that was switched to the 50th the measurement dropped to something like 120-140gig

The amount of traffic doesn't drop at all, it is jus the way you measure it. I think billing on average is good, but when you get billed on 95th by your upstream billing on average kills you.

No doubt , I know when you buy direct from a telco you get billed at the 95th and so billing by 50th for hosts that are in that situation causes a huge gap in revenue . I was stating the differences in the measurements and completely agree with you jay . But it doesn't change the fact that some of us are fortunate enough to have our provider suck up the costs or get reimbursed by their other clients to the point where they can give a few of us these great deals. Of course its not suited for those that buy small amounts of bandwidth but rather for those that use 4,5,10 mbp .

Originally posted by jayglate
The amount of traffic doesn't drop at all, it is jus the way you measure it.Exactly, and that's what has created confusion when we've discussed this here at WHT in the past. There's a common misperception that because the numbers are higher when you're paying under the 95th rule than when other methods are used, you're paying for "more bandwidth."

Maybe this is a decent analogy, at least I've been able to use it to help a couple of clients understand this: say you take a taxi across town. You go twenty miles, so the amount of roadway that was "transferred" under your wheels is twenty miles. You pay, say, 25 cents a mile so your bill is $5.

Now, imagine there's another kind of taxi, where you're charged not by how far you go but by how fast the cab goes! Every thirty seconds the car's speed is measured, and you have to pay ten cents for every mile per hour of your fastest speed -- or, say, 95% of that number. Your taxi travels the same ten mile route, but at one point it hits 60 miles per hour. You're billed only for 95% of that, or 57 mph. You pay $5.70. But if you'd hired a slower car, it would have gone no faster than 40, and you'd have paid only $3.80.

In each case the units you're measuring can be expressed with the same term: miles. But in the first case you only "used" 20 miles, in the second you used 57 and in the third 38. Sounds like you used more with the pay-by-speed taxi, but actually what you've measured are two different, though related, units.

But measuring at 50% in stead of 95% doesnt mean we get any slower than you do just becuase yours is measured different. For example at DI whether you use 95th or 50th you still get uncapped OC-12 fiber and use of the lines to the same capacity using the 50th as you do using the 95th. So we can still take that same cab ride and hit 60mph but not get charged for those peaks as much as you do on the 9th , this doesn't mean we don't pay for peaks , just not as much.

it may seem like a good solution in the short term. but traffic follows a curve (akin to a sine wave) throughout the day, and if you're only getting billed at the 50th percentile, _someone_ isn't paying for the size of the pipe. and that's a bad thing because it means that _someone_ isn't going to have significant scalability when you start approaching the limits of your link.

95th percentile is a good (and fair) billing method that gets users to pay what they are using but without penalizing them for the occasional spike in traffic.

people can save money and offer ridiculous (< $2) per mbit pricing by going with something other than 95th (or 90th) percentile billing, but in the end, it's not a long term viable way of doing things.

_all_ of the major tier-1 carriers bill this way. there's a reason why this is so ;-)

{ and don't get me started on the craziness of using
mrtg or cricket to do billing }

Raj Dutt

Originally posted by Jag
But measuring at 50% in stead of 95% doesnt mean we get any slower than you do just becuase yours is measured different. For example at DI whether you use 95th or 50th you still get uncapped OC-12 fiber and use of the lines to the same capacity using the 50th as you do using the 95th. So we can still take that same cab ride and hit 60mph but not get charged for those peaks as much as you do on the 9th , this doesn't mean we don't pay for peaks , just not as much.

Hasn't been a problem yet and given DI's reliability and scalability we are not the least bit worried. Cheers though

the only time i've seen it become a problem is when a provider gets too close to capacity (ie 80%). of course, they can make it up by charging a higher mbit price. can i ask what your per mbit price with DI is?

also, is DI Digital Island or Dialtone Internet?

Cheers. And good luck! :)

Raj


Originally posted by Jag
Hasn't been a problem yet and given DI's reliability and scalability we are not the least bit worried. Cheers though

sorry, DI = Dialtone Internet

Originally posted by nopzor
I agree with you completely.

95th percentile makes sense for those who value a stable and speedy network.

It is a fair way of paying for usage based on capacity.

Cheers,

Raj Dutt
Voxel dot Net, Inc.





Ditto. Quality bandwidth is almost always billed as 95th percentile.

"don't get me started on billing with mrtg or cricket"

Ok, you've got my attention -- how DO you bill for 95th without otherwise buying a $100,000 billing system from HP or someone like that?