CSE 561 Notes - October 21, 2002
Protocol
- Goals
- end-to-end bandwith
- (minimize lost packets)
- efficiency vs. predictability
- fairness
- host latency
- stability/predictability/Q.O.S.
- robust w.r.t. unforeseen topologies/workloads
- security
- (we'll ignore this for now)
- How do we allocate resources?
- Which brings us to...
TCP (duplex/sliding window)
- What happens when you blast the window?
- lost packets
- if routber buffer < window size
- if link bandwith < access link
- Therefore the technique is to:
- (1) Find out bottleneck bandwith
- slow start: exponential increase until packet loss observed
- (2) slow start again until you get to highest bandwith that works
- (3) then use AIMD (additive increase multiplicative decrease)
- each time there is multiplicative decrease, use slow start again
- A's throughput (xput) and B's xput will converge to most fair point.
- How to avoid the timeout persiod (see seond picture): fast retransmit
- if received 3 duplicate acks in a row:
- assume loss and retransmit
- (remember: TCP will awk a recieved packet with last in-order received packet)