Controlling High-Bandwidth Flows at the Congested Router

1
Controlling High-Bandwidth Flows at the Congested
Router

• Ratul Mahajan, Sally Floyd, David Wetherall
• ATT Center for Internet Research at ICSI (ACIRI)
• and ATT Labs Research

Presented by Yan Gao
2
Oscillation

• Paper used below method to reduce oscillation
• The reduction in the dropping probability is
  bounded by a max_decrease in one step
• The oscillation can also be caused by reaction
  delay and imprecision of using a flows packet
  drop history as an estimate of its arrival rate.
• Changes to the dropping probability are not make
  until a certain time period. How to set this
  certain time period?
• The paper did not explore the potential for
  oscillations in the average queue size, in
  particular for scenarios with large propagation
  delays and long-lived traffic.
• In practical, the average queue size varies
  significantly.

3
Mechanism

• It can be only applied for non-burst packet drop
  behavior.
• How about burst packet drop case?
• Detecting unresponsive flows is tricky in
  variable demand environments

4
Mechanism

• Since fast memory is not required for storing the
  drop history, does it affect the speed of online
  operation?
• This paper didnt show the precise state
  requirements and fairness tradeoffs under various
  traffic scenarios.
• RED-PDs test for unresponsiveness can have false
  positives and false negatives.
• What are the results of their evaluation?

5
Parameter setting

• Changes in the network might expose the
  importance of different parameters.
• The parameters have proved important in the past,
  so it should provide observations of their values
  on the Internet, and describe expected effects of
  other values.
• How could they set these parameters? Do they have
  some recommendation for different network? (such
  as K, M)

6
Evaluation

• The authors only did some very small level
  simulation by NS-2 (10Mbps), how about the
  behavior on large scale network (such as 10Gbps)
  ???
• The paper did not investigate performance for a
  range of link bandwidths for the congested link.
• Whether the approach will reduce totally
  bandwidth waste is still unknown.
• Bandwidth is wasted by delivering packets through
  the network that are dropped before reaching
  their ultimate destination
• There is no good way for responding to
  unresponsive flows.