Hello, We are working on a research project on interdomain traffic engineering and would like to quantitatively evaluate interdomain traffic engineering solutions in a realistic way. We have a BGP simulator that allows us to simulate networks with a few hundreds of AS and are adding several mechanisms to this simulator. To complement the simulator, we are looking for realistic cost models of transit services. Since we do not buy transit services, we have some difficulties finding this information. For the moment, our cost models are the following : - flat fee for a L Mbps link - volume based, y $ per Mbps (95% quantile) for a L Mbps link - burstable, flat fee for x Mbps on a L Mbps and z $ per Mbps above x We would like to know whether these three cost models accurately reflect the current pricing schemes of transit services and what are the most popular ones ? If you know other pricing schemes that should be taken into account, we would appreciate to have additional information. I heard of destination-based pricing schemes, but do not have enough details on them and don't know whether they are frequently used. You can either answer directly to us (Olivier.Bonaventure@info.fundp.ac.be and suh@info.ucl.ac.be) and we will summarize or directly to the list. Thanks a lot for your help, Olivier Bonaventure --- http://www.infonet.fundp.ac.be
- flat fee for a L Mbps link
Also known as 'fractional' or 'tiered.' $x for y mb/s, and it is rate-limited.
- volume based, y $ per Mbps (95% quantile) for a L Mbps link - burstable, flat fee for x Mbps on a L Mbps and z $ per Mbps above x
These two are essentially the same. You do have three variations of usage-based, however: a) vth percentile: $x per y zzzbits/sec, with a t committment. Occasionally, any usage over t has a different price. b) 'Average usage', which is is the same as A, but using an averaging measurement system, rather than a percentile system (50th percentile is NOT the same as, or even relevant to, average). c) counting bytes: $x per y bytes. -- Alex Rubenstein, AR97, K2AHR, alex@nac.net, latency, Al Reuben -- -- Net Access Corporation, 800-NET-ME-36, http://www.nac.net --
On Mon, Sep 23, 2002 at 11:26:22AM -0400, Alex Rubenstein <alex@nac.net> wrote:
- flat fee for a L Mbps link
Also known as 'fractional' or 'tiered.' $x for y mb/s, and it is rate-limited.
- volume based, y $ per Mbps (95% quantile) for a L Mbps link - burstable, flat fee for x Mbps on a L Mbps and z $ per Mbps above x
These two are essentially the same. You do have three variations of usage-based, however:
a) vth percentile: $x per y zzzbits/sec, with a t committment. Occasionally, any usage over t has a different price.
In my somewhat limited experience, "t" commitment is usually at least 10% of wire speed. Though if transit is coming off low-cost LAN ports in data center environments, sales folks will probably approve lower commits if pressured. Also, some large ISP's have a policy that you must buy the whole pipe unmetered if your commit is >50% pipe speed. And there are at least 4 ways of computing 95th percentile, though I'm sure there've already been threads on this. Cheers, -Lane
b) 'Average usage', which is is the same as A, but using an averaging measurement system, rather than a percentile system (50th percentile is NOT the same as, or even relevant to, average).
c) counting bytes: $x per y bytes.
-- Alex Rubenstein, AR97, K2AHR, alex@nac.net, latency, Al Reuben -- -- Net Access Corporation, 800-NET-ME-36, http://www.nac.net --
On Mon, Sep 23, 2002 at 12:50:17PM -0700, Lane Patterson wrote:
Also, some large ISP's have a policy that you must buy the whole pipe unmetered if your commit is >50% pipe speed.
Never heard that one, but conversly most ISPs have a minimum commit for "big expensive ports". For example, 1 meg commits on FastE ports are usually fine because almost nobody still has ports that are only 10Mbit Ethernet. But noone in their right mind will give GigE ports to 10Mbit committers, for potential abuse reasons and port cost reasons at the very least.
And there are at least 4 ways of computing 95th percentile, though I'm sure there've already been threads on this.
There is only one way, anyone else is computing "something else" that they just happen to bill with. But this sounds like a subject for the NANOG FAQ. :) -- Richard A Steenbergen <ras@e-gerbil.net> http://www.e-gerbil.net/ras PGP Key ID: 0x138EA177 (67 29 D7 BC E8 18 3E DA B2 46 B3 D8 14 36 FE B6)
On Mon, 23 Sep 2002, Richard A Steenbergen wrote:
On Mon, Sep 23, 2002 at 12:50:17PM -0700, Lane Patterson wrote:
Also, some large ISP's have a policy that you must buy the whole pipe unmetered if your commit is >50% pipe speed.
Never heard that one, but conversly most ISPs have a minimum commit for
I have, but not 50%.. usually around 75% and the way I've seen it is you pay for the 75% and that gets you the full pipe unmetered. ie the extra 25% is FOC but its not really that usable anyhow as you know.. Steve
"big expensive ports". For example, 1 meg commits on FastE ports are usually fine because almost nobody still has ports that are only 10Mbit Ethernet. But noone in their right mind will give GigE ports to 10Mbit committers, for potential abuse reasons and port cost reasons at the very least.
And there are at least 4 ways of computing 95th percentile, though I'm sure there've already been threads on this.
There is only one way, anyone else is computing "something else" that they just happen to bill with. But this sounds like a subject for the NANOG FAQ. :)
On Mon, Sep 23, 2002 at 04:07:47PM -0400, Richard A Steenbergen wrote:
On Mon, Sep 23, 2002 at 12:50:17PM -0700, Lane Patterson wrote:
And there are at least 4 ways of computing 95th percentile, though I'm sure there've already been threads on this.
There is only one way, anyone else is computing "something else" that they just happen to bill with. But this sounds like a subject for the NANOG FAQ. :)
I think the problem is not that there are multiple definitions of how to calculate the 95th percentile of a sample population, but that different peoples' sample populations are constructed in different ways. I have seen billing based on: max(95th(to_cust), 95th(from_cust)) 95th(max(to_cust, from_cust)) 95th(to_cust + from_cust) 95th(to_cust) all referred to by different people as "95th percentile billing". There are also many different ways of obtaining values for "to_cust" and "from_cust", just to keep things interesting. I would be interested to find out how many customers of services billed by "95th percentile" do their own measurements and compare them with the bill. I suspect the number is not large. Maybe that's why most providers don't find it necessary to spell out exactly what calculations they are doing in order to arrive at a monthly figure with a dollar sign in front of it. Joe
On Mon, Sep 23, 2002 at 05:28:15PM -0400, Joe Abley wrote:
I think the problem is not that there are multiple definitions of how to calculate the 95th percentile of a sample population, but that different peoples' sample populations are constructed in different ways.
I have seen billing based on:
max(95th(to_cust), 95th(from_cust)) 95th(max(to_cust, from_cust)) 95th(to_cust + from_cust) 95th(to_cust)
There are really two things we're talking about, the mathmatical function for "95th percentile", which is quite simply take the highest sample after removing the top 5% of samples, and the "95th percentile billing system", which can only be defined as max(95th(to), 95th(from)). Anything else is a varient which may or may not be useful as a billing metric, but it is NOT "the 95th percentile billing system", it is "a billing system which happens to use the 95th percentile function".
I would be interested to find out how many customers of services billed by "95th percentile" do their own measurements and compare them with the bill. I suspect the number is not large.
I suspect most customers don't even fully understand 95th percentile. I still see many people thinking its the average after the 5% is removed, infact if you're a good you can convince sales people at "some major ISPs" to write it that way in the contract.
Maybe that's why most providers don't find it necessary to spell out exactly what calculations they are doing in order to arrive at a monthly figure with a dollar sign in front of it.
The stories I've heard from the few people who do their own measurements for billing confirmation usually involve some significant over (or sometimes under) billing. As they put it, "We get a bill for 750Mbit, we send them the money for the 670Mbit we actually used, and they're happy." -- Richard A Steenbergen <ras@e-gerbil.net> http://www.e-gerbil.net/ras PGP Key ID: 0x138EA177 (67 29 D7 BC E8 18 3E DA B2 46 B3 D8 14 36 FE B6)
Also, some large ISP's have a policy that you must buy the whole pipe unmetered if your commit is >50% pipe speed.
The beauty of current economic situation is that should that ever be someone to claim that one's answer can always be "Why dont you pass me to your manager" at which point the problem gets resolved into the buyers favour. It does sometimes require pointing out the the manager that their company had been missing its own revenue guidance.
And there are at least 4 ways of computing 95th percentile, though I'm sure there've already been threads on this.
There is only one way of computing 95th percentile. Since it is math, it is very black and white. There are multiple ways of defining what are the inputs of 95th percentile. Alex
participants (7)
-
Alex Rubenstein
-
alex@yuriev.com
-
Joe Abley
-
Lane Patterson
-
Olivier Bonaventure
-
Richard A Steenbergen
-
Stephen J. Wilcox