The Macroscopic behavior of the TCP Congestion Avoidance

The Macroscopic behavior of the TCP Congestion Avoidance Algorithm

Introduction: • Analysis of a performance model for the TCP Congestion Avoidance Algorithm • Verification of the model through both simulation and live Internet measurements

Assumptions: • TCP Congestion Avoidance Algorithm in steady-state • Light to moderate packet loss • Additive increase/multiplicative decrease for dealing with congestion • Multiple losses within 1 RTT treated as 1 congestion signal • Random packet loss at constant probability p

TCP window evolution under periodic loss

Simple derivation: • W/2<Win<W • If (ACK each segment) each cycle must be W/2 RRT • Total data delivered ~ Area on the graph (3/8)*W 2 = 1/p • Solving for W: W=√ 8/3 p

We are interested in BW: • BW = (data per cycle) / (time per cycle) = (MSS* (3/8) * W 2 ) / (RTT * W/2) = (MSS / p) / (RTT * √ 2/3 p ) • BW = ( MSS / RTT ) * (C / √p ) • C is a constant “around” 1

Where it does not fit: • advertised window too small cwnd has no effect • Sender has no data • Time consumed by timeouts not modeled • Go-back-N strategies (window consumed by needless retransmissions) • other window opening strategies (Vegas) • Short connections ( not steady-state )

Two examples: • Queueless random packet loss (constant delay, high BW link) • Environments with queuing (drop-tail bottleneck link)

Measured vs. Estimated BW

Window vs. Loss

Estimated Window vs. Loss I

Estimated Window vs. Loss II

Fitting the Slope • Window vs. loss data ~ linear, but… • BW = ( MSS / RTT ) * C * pk , k~1/2

Analysis of the Increase and Decrease Algorithms for Congestion Avoidance in Computer Networks

Introduction: • Both congestion avoidance and congestion control mechanisms are basically resource management problems • Key component of a congestion avoidance scheme is the algorithm • Metrics: efficiency, fairness, convergence time, size of oscilations

Network performance = f(load)

• congestion avoidance – operate at the knee • congestion control – left of the cliff • One can limit number of packets (window mechanism) or rate (bits or packets per second)

Binary feedback scheme: • Decentralized decision-making algorithm • Resource sends single bit, congestion experience bit (1 - overloaded, 0 underloaded ) • Users then adjust their load by increase/decrease algorithm

Algorithm’s control function : • • • xi(t) - user’s load ui(t) - user’s control y(t) - binary feedback ui(t) = f (xi(t), y(t) ) xi(t+1) = xi(t) + f (xi(t), y(t) )

Focus on linear control function f • xi(t+1)= a + b xi(t) • 4 combinations: multiplicative/additive + increase/decrease

Criteria for selecting controls: • Efficiency of the resource usage X(t) – closeness of the total load to the knee • Fairness estimated by the maxmin criterion, xi(t)= xk(t), I, k share the same bottleneck F(x)=(Σxi )2 / (Σ xi 2 ) • Distributedness –min amount of feedback in the system • …

Convergence – speed/time with which the system approaches the goal state

A Web Server’s View of the Transport Layer

• Observations of traffic to and from a particular WWW server • Data collection from Apache logs and tcpdump

Selective Acknowledgments

Round Trip Times

Larger Initial cwnd

Conclusions: • SACK based TCPs should be suported by servers • 85% RTTs are between 15 ms and 500 ms • using larger initial cwnd does not drastically increase no. of connections experiencing loss in the first burst of data