Distribution and allocation

Every entity has an assigned number of points inversely proportional to that entity's priority. The global rate limit is divided by the total number of points over all entities, and this is the rate allotment per point. If every entity performed I/O with a rate equal to its number of points multiplied by the per point allotment, the total rate over all entities would be at the rate limit and every entity would perform I/O at a rate proportional to their assigned priority. Since Trickle is performing bandwidth shaping, most often the entities has the ability to exceed the transfer rates that they are assigned by the scheduler. The entities only very seldomly behave in any predictable manner: Their data transfer rates may be bursty, they may have consistent transfer rates lower than their alloted rates, or they might be idle. At the same time, the scheduler needs to make sure that the entities in aggregate may transfer data at a rate capped only by the total rate limit: Trickle should never hinder its client applications from fully utilizing the bandwidth allocated to them.

Statistics structures as well as limits are kept independently per-direction and thus Trickle is fully asymmetric. Furthermore, it is worthy to note that Trickle makes scheduling decisions based on a windowed average. Dealing with instantaneous bursts is discussed later.