Congestion Control for High Bandwidth-Delay Product Networks D. Katabi (MIT), M. Handley (UCL), C. Rohrs (MIT) – SIGCOMM’02 Presented by Cheng Huang

Basics of TCP Congestion Control Bandwidth-delay product  Capacity of the “pipe” between a TCP sender and a TCP receiver Congestion window (cwnd)  sender’s estimation of the capacity Additive Increase and Multiplicative Decrease (AIMD) algorithm  no loss:cwnd = cwnd + s  loss: cwnd = cwnd – cwnd/2

Motivations Inadequacy of TCP, as bandwidth-delay product increases  Prone to instability regardless of AQM schemes  Inefficient Fairness concern  TCP tends to bias against long RTT flows Satellite links, wireless links, etc.

Design Rationale NOT an end-to-end approach Using precise congestion signaling Decoupling efficiency and fairness control

Features of XCP (eXplicit Control Protocol) Maintains high utilization, small queues, and almost no drops, as bandwidth/delay increases  drop: less than one in a million packets Maintains good performance in dynamic environment (with many short web-like flows) No bias against long RTT flows

XCP – Sender/Receiver ’ s Role Sender  Fill the congestion header  Update cwnd = max(cwnd + H_feedback, s) Receiver  Copy H_feedback to ACK

XCP – Router ’ s Role Control Interval Estimation  Average RTT Efficiency Control  Maximize link utilization Fairness Control  Achieve fairness among individual flows

Control Interval Estimation Estimation requirement  Core stateless  Average over flows (not over packets)  e.g. two flows have RTTs of 80 ms and 40 ms and the same cwnd = 10 packets, then average RTT over packets is: RTT avg = (80*10+40*20)/(10+20) = (ms) Instead, average RTT over flows is: RTT avg = (80*80*10+40*40*20)/(80*10+40*20) = 60 (ms)

Control Interval Estimation (2) Weight of each packet  w i = H_rtt i * (s i / H_cwnd i ) Average RTT  sum(w i * H_rtt i ) / sum(w i ) Average cwnd  sum(w i * H_cwnd i ) / sum(w i )

Efficiency Controller (EC) Aggragate feedback (total H_feedback)  alpha, beta: constant value  d: control interval (average RTT)  S: spare bandwidth  Q: persistent queue size Stability requirement determines alpha = 0.4; beta =  Independent of delay, capacity and number of flows

Fairness Controller (FC) Achieve fairness via AIMD algorithm  phi > 0, equal throughput increment of all flows  phi < 0, throughput decrement proportional to its current throughput Positive feedback  (1/w i ) * (p i /H_rtt i ) = C 1 (constant value)  sum(p i /H_rtt i ) = phi/d Negative feedback  (1/w i ) * (n i /H_rtt i ) = C 2 * H_cwnd i /H_rtt i  sum(n i /H_rtt i ) = phi/d

Fairness Controller (FC) (2) Bandwidth shuffling  h = max(0, gamma*y - |phi|) gamma = 0.1 y: input traffic

Performance Evaluation Simulation topology I

Performance Evaluation (2) Simulation topology II

The dynamics of XCP (I)

The Dynamics of XCP (II)

Differential Bandwidth Allocation Replace FC  phi > 0: allocate throughput increment according to flows’ prices

Gradual Deployment – A TCP-friendly XCP Separate queues to distinguish TCP and XCP traffics Calculate average cwnd of TCP traffics by Update weights to make TCP and XCP fair

Conclusion XCP provides a theoretically sound, yet effective approach to congestion control. It remains excellent performance, independent of link capacity, delay and number of flows.