1. Performance Guarantees Introduction –by asking sources about flow behavior it is possible to construct networks that could guarantee performance for individual flows –traffic management involves queue management
2. Statistical vs Deterministic Deterministic guarantees : all data will arrive within a delay D Statistical guarantees : some percentage of data (very large) will arrive within the delay D
2.1 What is guaranteed ? worst case loss rate, bandwidth, queueing delay Effective loss rate : arrive to late or queue overflow Size of queue is important : - size and delay are related - if too small : overflow
3.Statistical multiplexing term for : taking traffic from a number of sources and forwarding it out the same output line rely on : law of large numbers the bandwidth is constant but does not state what the required aggregate bandwidth will be
is unguaranteed traffic purely random? for telephone networks is random data traffic is correlated even from a very large number of sources data network traffic is self-similar or fractal
3.1 Makrucki’s Proposal scheme for management of guaranteed flows in ATM networks if a source exceed its rate turn on CLP cells with CLP on may be dropped simple
3.2 Performance Bounds packet train effects becomes stronger as a flow path gets longer TCP acknowledgements can get compressed cells from the same connection are moved closer
3.3 How to deal with worst case? Constrain the sources, so that it is harder to form bunches Remix the traffic periodically Statistical guarantees expressed in term of connections with infinite lifetimes Used in backbones
4. Weighted fair Queuing Maintain a queue for each source and forward the packets in round-robin order - ignores packet length - sensitive to the patterns of packet arrivals bit by bit round robin weighted fair queuing
4.1 Performance Expensive to implement - maps each packet to its flow - maintain a queue
5. Jitter Control Schemes In weighted fair queueing packet will arrive no later than t+D t+x x can vary by a value called jitter
5.1 Stop and Go Queuing Input frames Output frames All frames have T bits
Delay D = p + Δ+ δ p - propagation delay Δ - fixed delay δ - small variance h T < Δ < 2hT, h - path length -T < δ < T Delay is bounded by p+Δ+δ
5.2 Jitter EDD Jitter Earliest Due Date When flow is established : - each hop is given a queuing budget b - each hop has a jitter bound j When packet is transmitted records the difference σ = t-b At next hop σ is added to b
6. Statistical multiplexing revisited Weighted fair queueing, Stop and Go, jitter EDD : guarantees are very long FIFO+ : incorporate low delays of statistical multiplexing with delay guarantees
6.1 FIFO + Limit delay accumulation - Hops groups flows into classes - Difference between the queueing delay and average queueing delay of class