Myths About Congestion Management in High-Speed Networks

One-line summary: Provides substance on both sides of several religious arguments on congestion control, including open-loop vs. closed-loop, one scheme vs. many, router vs. source, backpressure vs. explicit feedback; concludes that heterogeneity of traffic and network types is making congestion worse, and "one size fits all" will never work.

Overview/Main Points

• Previous myths exploded: Congestion will go away when buffers get bigger, WANs get faster, CPU's get faster, or all of the above.
• Economic arguments for fast WANs:
  • high-speed means more sharing, therefore better cost amortization. Analogy: highways (hi speed) connect cities (low speed), even though most traffic is local.
  • Today's data networks carry voice traffic as packets; voice networks carry data as coded voice signals. This separation cannot persist because it is not economical.
• Rate vs. window based congestion control: Rate-based should become more popular because memory is no longer the bottleneck, because window based flow control is more susceptible to burstiness, and because much high-speed traffic will require rate rather than window guarantees. (One possibility: rate-based with fallback to a large window size, to be used in case of severe congestion.)

  What	Window-based	Rate-based
  Control	Window	n packets every T seconds (most models incorrectly collapse these to single param L=n/T)
  Eff. rate	Window/RTT	n/T
  Required if	Memory is bottleneck	CPU, link, or rate-based devices are bottleneck
  Max Q length	determined by window	unbounded
  Burstiness	yes	not at source
  Controlled	end-to-end, hop-by-hop, or both	hop-by-hop
  Network layer	Connectionless or connection-oriented	connection-oriented

• Router- vs. source-based:
  • Arguments against source based: significant delays; sources may not cooperate; feedback may require additional packets to be injected (explicit feedback notification); fairness can't be achieved.
  • Argument against router-based: adds instructions to critical path of packet forwarding; congestion continues until sources reduce their traffic anyway.
  • Summary: use router-based to ensure fairness and for short-duration overload; source-based (and at higher network layers) for longer overloads.
  • Backpressure is a datalink level hop-by-hop on/off flow control that works well in small-diameter networks for short durations. Unfair in that traffic following uncongested path can be adversely affected.
• Reservation vs. walk-in: Reservations good for steady, predictable traffic; walk-in good for shorter, bursty traffic.
• One scheme vs. many: need one, since congestion control deals with shared resources. Rule of thumb: the longer the duration of congestion, the higher the layer in which congestion control should be implemented.

Relevance

Flaws

• Most arguments, though intuitively satisfying, are qualitative rather than quantitative, though there are references.
• Not a prescriptive paper; given all this information, and the existence of the heterogeneous wide-area Internet, what do we do?