This paper looks at a model of reducing peak-rate load by incentivising users to move from peak rate slots to off-peak time periods. It has its roots in their HotNets 2008 paper “Good things come to those who (can) wait”. (Users are granted bandwidth in the off-peak for good behaviour in the on-peak.)
Data: The dataset is from a large transit provider. 12 million ADSL users uplink and downlink volumes over 5 minute intervals for several weeks in 2008. The provider connects to over 200 other networks. WIDE network data (Japan link) is also used. They classify the traffic by application – unfortunately the two scenarios lead to estimates of P2P traffic varying as 12–22% of all traffic throughout the day (pessimistic assumptions about classification or 74–88% throughout the day (optimistic assumptions).
Model: The model is of a user's traffic as a vector over “slots” in time. The ISP “bids” to make users move traffic to different slots. The model also incorporates a “peak hour” which is the set of busiest slots (defined by a link utilisation). A threshold is set which is the maximum desirable utilisation. The model is then to rearrange the users’ “elastic traffic” while providing the extra bandwidth and keeping the peak bandwidth below . ISPs can be “omniscient” (offering different incentives to different users) or “oblivious” (offering the same incentive to all users). The incentive offered to user in the omniscient case is an off peak bandwidth of (this is the total amount of extra traffic which can be sent in the off-peak) where is the amount of traffic shifted away from the peak hour and is the increased bandwidth factor. (Note, it would seem that, if not carefully handled, this could create an incentive for a user to make spurious downloads to increase their capacity off-peak.) Users are split into “all-or-none” and “fractional” users. The former shift all elastic traffic or none of it, the latter may shift a fraction. The authors sketch a proof that the “fractional users” and “omniscient ISP” problem is solvable in polynomial time.
The incentives is modelled as an ISP “bid” (by offering a for the user to shift their elastic traffic. User “greediness” is modelled as resistance to move (or the amount of increase required to make them move). This is distributed between users, in one model using a pareto distribution and in another as a flat distribution.
The modelling proceeds by picking a maximum utilisation for a link and looking at the amount of traffic expansion that the ISP has to provide in the off-peak to incentivise the users to make this move.
Page generated 2014-01-07, by jemdoc.