On the 95 percentile billing method

Authors: 
Xenofontas Dimitropoulos, Paul Hurley, Andreas Kind and Marc Ph. Stoecklin
Published: 
Passive and Active Measurement Conference
Year: 
2009

This paper describes the commonly used 95-percentile billing method which is often used by ISPs to bill other ISPs. The 95-percentile billing method is as follows:

  • Set a billing rate $y per Mbps (Megabit per second).

  • Take a month of traffic counts for the entity you wish to bill.

  • Split the traffic into equal sized time periods of length T (often 5 minutes).

  • Calculate the mean rate in Mbps for each time period.

  • Find the rate x_{95} Mbps of the time period which is the 95th percentile – that is 5% of time periods have more traffic and 95% have less traffic.

  • The total bill for that month for that entity is x_{95} y.

  • Sometimes it is the case that x_{95} is calculated separately for inbound and outbound and the maximum taken.

  • Usually it is not the actual traffic which is measured (by packet) but an approximation reconstructed from netflow data.

The aim of the paper is to look at how choice of the time period t and how aggregation of billed entities affects the amount billed. The data sets are

  1. Netflow trace from web host with 46 websites (27 days in April 2008) – lowest volume websites are stripped from analysis

  2. tcpdump trace from medium volume website on this host (30 days in July 2007)

  3. Medium sized enterprise campus network (transit services from commercial and academic ISP) this is a tcp dump for 63 days (March 2008)

Findings about window size are:

  1. Usually making the window size t smaller will make x_{95} larger

  2. In atypical cases this can be reversed (larger windows make x_{95} larger).

  3. However, 11 from 34 sites were atypical so typical is not very typical.

  4. Higher mean traffic levels tend to make the dependence on t weaker.

  5. This is related to the self-similarity of web traffic. For self-similar traffic he 95th percentile should decrease polynomially as the aggregation window t increases.

Netflow data is often used for billing individual sites. Volumes and lifetimes of flows are collected and it is assumed that the volument of a flow is smoothed across its lifetime. In some cases actually looking at the traffic in terms of packets not flows can make a difference because of this approximation. In most traces there was little difference between exact packet accounting and approximate flow accounting. However, the second month of campus data showed a signficant (30%) difference. This is because this data set shows a large number of long-lived flows which ‘‘spike" periodically shifting up the 95-percentile.

The main conclusion here is that ISPs should use a standardised method to calculate 95-percentile to allow fair and easy comparison of billing across possible providers.

Show bibtex
Page generated 2014-01-07, by jemdoc.