[ NNSquad ] Costing P2P and other ISP data streams

["Neil Davies" <semanticphilosopher@googlemail.com>]

P2P and its price/cost has been discussed several times on this list in the recent past. I'm submitting this post as an example of a costing approach. Although it is based on UK pricing, the modelling approach has wider application.

Some notes:

1. this is about costs not prices - I am far more a mathematician than a marketeer - these numbers reflect the underlying cost of deliver services, not the price that they would be retailed.
2. there are a number of assumptions that have been made - caveat lector - see[1] for some of the important caveats.
   
3. assuming that the ISP has the wherewithal and  mechanisms to implement the policies stated

Preamble

There are (in UK) four components to a ISP's data movement costs:

• the end-user tail rental
• the cost of the connection(s) to wholesale provider's access network
• the cost of the (aggregated) bandwidth into that  wholesale provider's network
• the cost of the  interconnection and bandwidth to wider internet

I am going to ignore the end-user tail cost - it is a fixed per monthly charge independent of end-user usage - as such a constant in this model.

Aggregated bandwidth is charged at peak[2] - you pay for the whole month for the highest bandwidth[3] consumed in a measurement period. This peak usage  model is not unusual in the electricity industry for large consumers - just like you can't store transmission bits for later use you can't store electrical power.

Although the headline figure for aggregated bandwidth may appear to lie in the $150-$200 per mbit  per second month[4], these prices don't incorporate the other access costs. A more conventional cost, if you were paying only on the basis of peak bandwidth would be $300-$400 per mbit per second month. To make the arithmetic easier I am going to assume the $400 figure, or stated another way $0.40 per kbit per second month.
 
Cost Model

Assume that your ISP can fix the maximum instantaneous rate it sends data to the wholesale network (if it can't it has other risks it has to manage). Their logic for setting that peak rate would follow the following sort of lines:

There are peak periods of demand (middle of day for business oriented ISP, early evening for Retail ISP). Now assume the ISP is going to "guarantee" certain services an assured bandwidth (ignoring the issue of the quality of that bandwidth for the moment). The cost to ISP (the price it must charge so as not to make a loss) is $0.40 per kbit per second for that service.

Using this they can establish their costs for certain services - e.g the per  Erlang cost of VoIP, Video telephony - these are services that have discrete units of consumption, as well as the cost of other services such as VoD, interactive terminal services etc.

Most of the discussion on this list recently has been about P2P.  Before we cost this a little more introduction is needed

P2P &  Scavenger 

From the neutrality standpoint lets not call this P2P traffic -  that is too emotive about its content. Let's call this scavenger traffic - it is bulk transfer traffic that the end user has no particular time constraints over getting transmitted (it might be backup traffic, it could be microsoft patch traffic, etc). The important property is that it should get there sometime, the movement of the traffic should have no effect on the end-user's other network traffic (i.e VoIP still works, the normal web browsing continues smoothly). It is a background, non-intrusive, data movement service.

Also assume that the end-user is working with the ISP to identify this traffic (e.g arrange for their bit torrent to be on particular ports)- they want this service because it doesn't get in their way when they do other things on the net and it saves them money... read on.

Let us also assume that there are limits on the per-monthly usage - I'm going to pick 10GBytes per month[5] - you can scale it up or down from there as you wish[6].

Delivering 10GBytes (of user data) evenly over a month requires (approx) a dedicated 32kbits/sec.

The cost of reserving that 32kbit/sec for the peak periods is $12.80.

This would seem a bit steep, if the end user were to accept a lower minimum rate when the rest of the network is in saturation, say 4 kbits per second then the ISP cost would be $1.60. This would not stop the end-user moving data at other periods at faster rates (there are statistical multiplexing assumptions here) - there has to be the appropriate mix of traffic for the ISP to recover their peak costs across all their combined services. These conditions are not that onerous - but what they do say is that for cheap scavenger traffic the ISP needs a good mix of other services it can charge for to cover the peak bandwidth costs.

This seems to imply that an ISP incremental cost model (assuming end-user tail costs already covered)  for a "scavenger service" with a per monthly limit of 10Gbytes, minimum rate 4kbits/sec (could set a max if wanted say 1mbit/sec[7]) could be  $1.50 per month and a charge of $0.0001 per MByte transferred. 

The key point here is the minimum assured rate drives the cost, given the appropriate traffic mix, - the $0.0001 per MByte is a nominal cost - nothing is ever free, can be low cost - but there needs to be a usage charge. 

Modeling a scavenger service is rather different than other services - peer-to-peer networking basically means that the demand is effectively unbounded; The model needs to reflect his 100% duty cycle requirement - if its there it will use it!

This is more than just idle speculation on my part - we've got access to usage statistics and this sort of approach seems make sense, even on the small scale - all you seem to need is a large enough user population to permit the statistical properties of samples to start holding.

Hope that this helps - looking forward to the discussion

Neil

[1] There is a lack of precision in the description that wholesalers' supply in how bandwidth is measured, it is not necessarily measured in say as IP packet size, it may be measured as those packets encoded within the underlying data carrier (i.e the number of ATM cells), it may be a uni-directional (downstream) or a bi-directional measure. Bandwidth is calculated by calculating volumes transmitted within a sampling period; so it is not really an instantaneous peak. There appears to be a trend to reduce the fixed costs (say for the connection to the wholesaler's access network) offset by a corresponding increase in bandwidth charges. Wholesalers typically require you to contract to a particular monthly peak, they will allow you to consume more than that peak. Such excesses attract an additional surcharge ranging from 25% to 75%.
[2] Some charge at the 95th percentile of measurements over the billing period. Using peak here as it is easier to grasp. 
[3] Its not really a  bandwidth - it is a volume in a period. Period is, typically, 15 minutes - So 1mbit/sec is really 54 Gbits measured over 15 minutes. This becomes a key difference when assuring loss and delay characteristics to streams.
[4] An exchange rate £1 = $2 is used throughout and broadly rounded.
[5] In actual practice it might be 10GBytes over a rolling 30 day period (you don't want to reset all the counters to zero at the start of a month - mayhem will ensue); additionally you might want additional limits over other periods e.g. 10GBytes per 30 days implies no more than 5 GBytes over any 7 days or  2.5Gbytes over any 24 hours. 
[6] Beware linear scaling; the ISP's capacity to the network is a finite resource - fundamentally the law of supply and demand states that the cost function for a commodity should be monotonically increasing as the demand approaches the supply. Expect to charge/pay more if you demand a larger fraction of the network capacity.
[7] This would permit the end user to reach their 24 hour limit (see [5]) in  6 hours