Congestion Control for High Bandwidth-Delay Product Networks

Introduction

• math shows that regardless of queueing scheme –> higher bw-delay product causes TCP to oscillate between bandwidth
• all previous attempts are eventually oscillatory at some point
• additive policy slow to adapt
• addition of bandwidth or flows slow to adapt
• generalize the ECN concepts
  • split utilization control from fairness control
  • control state within packets

Design

• think about congestion control w/o backcompat concerns…
• packet loss is not a good congestion signal
• aggressiveness of changes should be adjust based on a feedback loop
• robustness of congestion should be independent of unknown parameters (e.g. number of flows)

Protocol

• join design of end-systems and routers
• senders maintain congestion window and rtt
  • communicate to routers via congestion header
• congestion header has current sender values, and one feedback slot for pos/neg number
• XCP sender
  • sender initializes the feedback to desired window increase
  • whenever ack arrives, positive feedback increases sender congestion window, negative reduces it
• XCP receiver
  • router must compute feedback to cause system to converge to optimal efficiency and min-max fairness
  • router has efficiency controller, fairness controller
  • make control decision every RTT
    • per-link estimation control timer to calculate estimate
• efficiency controller
  • maximize link utilization while minimizing drop rate and persistent queues
  • looks at aggregate traffic, does not care about fairness
• fairness controller
  • apportion feedback to individual packets to achieve fairness
  • similar principle as TCP, AIMD
• MIMD in efficiency controller with feedback ~1 RTT to maximize bandwidth utilization
• estimation errors are ok because estimations change frequently (1 RTT)

Performance

• Across different bottleneck link speeds…
  • XCP nearly always uses all bandwidth
  • XCP bottleneck router queues increase linearly
  • virtually no packet drops since linear increase in queue size
• across varying RTT propagation delays
  • high utilization
  • low queue size (not necessarily linearly increasing)
  • no packet drops
• across varying number of flows
  • bottleneck utilization decreases and at least matches for large numbers of flows
  • queue length linearly increases
  • no drops
• security
  • all nodes need to comply with feedback in order to work properly…
  • policing agents? -