Statistical/soft/predictive QoS
Goal: bound the tail of the measure distribution
Not a good idea to use worst-case delay bounds since very few packets (or none!) will actually experience this worst-case delay
Computing statistical bounds (e.g., using effective bandwidths) is usually approximate and often conservative
FIFO+ attempts to reduce worst-case delay and jitter using minimal isolation (and maximal statistical gain)
At each router, a packet is assigned lower priority if it left previous routers ahead of measured average delay, and higher priority if behind average delay