---

Active Queue Management: Random Early Detection/Drop

---

• Network Problem to be solved:
  • Synchronized packet losses among different (TCP) flows

  ---

  ---

• Protocol level of solution
  • Application level solution:

    Each computer runs an application program Different applications must become aware that they have a common bottle neck router Solution requires distributed awareness Solution at the application level is not practical...

    ---

  • Transport level solution:

    Each computer contains a TCP module Different TCP modules must must become aware that they have a common bottle neck router Solution requires distributed awareness Solution at the application level is not practical...

    ---

  • Network (IP) level solution:

    Each router contains a IP module Synchronized TCP flows is caused by synchronized packet drops from different flows The solution is to prevent or reduce the likelihood that packets from different flows are dropped synchronized

    ---

    ---

  • $64,000 question:

    How to prevent routers from synchronized packet drops from different flows during congestion.

  ---

  ---

  ---

  ---

• The General Random Early Detection Method
  • The general Early Random Detection method:

    specify a dropping policy does not provide details on how to implement the policy

    ---

  • Overview general RED:

    General RED algorithm:

    compute average queue length; if ( avg. queue length < minthreshold ) { admit new packet to output queue; } else ( minthreshold < avg. queue length < maxthreshold ) { mark new packet with probability Prob(queue length); } else { always mark new packet; }

  • The RED algorithm can be found on Page 7 of the paper "Random Early Detection Gateways for Congestion Avoidance"

    ---

  • The general RED algorithm leaves the following questions open (i.e., undefined):

    How do you compute the average queue length ? How do you calculate the drop probability Prob(queue length) ?

  ---

  ---

  ---

  ---

• Marking packets
  • Ideal implementation of Marking:

    Offending packets are marked with a bit The markingbit is returned by the destination to the sender When the sender receives a marking bit, the sender reduced the CWND

    Advantage:

    Packets are not dropped

    Disadvantage:

    Modifications are required to both the sending and the receiving TCP modules Also, the modification must be made to all existing TCP software (Or else, it will not work)

    ---

    ---

  • Reality:

    Offending packets are dropped A dropped packet will cause TCP (Reno) to reduce its CWND

  ---

  ---

  ---

  ---

• The realized Random Early Detection Method
  • Floyd and Jacobson proposed the following methods to compute:

    average queue length : use exponetial weighted moving average drop probability: this function

  ---

  ---

• Exponent weighted moving average
  • Exponential weighted moving average: weights newer values more heavily, old values weighted exponetially less

    Input numbers: q1, q2, q3, ....            Weight: w          Exponential weighted moving averages: average of (q1):                     w × q1 average of (q1, q2):               (1 − w) × w × q1 + w × q2 average of (q1, q2, q3):         (1 − w)2 × w × q1 + (1 − w) × w × q2 + w × q3 average of (q1, q2, q3, q4):    (1 − w)3 × w × q1 + (1 − w)2 × w × q2 + (1 − w) × w × q3 + w × q4

  ---

  ---

  ---

• Average queue length
  • Computing Exponentially weighted moving average:

    Meaning of the variables: curr_time = current time q_idle_time = the time when the queue last became empty wq = weight (a constant between 0 and 1) q = current queue length; if ( q > 0 ) { avg = ( 1 - wq ) * avg + wq * q; // Compute moving average } else { // Idle !!! x = f( curr_time - q_idle_time ); avg = (1 - wq)x * avg; // Because: wq * q = 0 // Note: weight is increased // if interval x is larger q_idle_time = curr_time; // Paper has a bug }

    Notes:

    The statement avg = ( 1 - wq ) avg + wq * q            can compute exponetial decay for older terms

    ---

    Example: Suppose avg = 0 and the first 3 queue length measures are q₁, q₂, and q₃

    Initially: avg = 0 First queue length measure: q1 avg = ( 1 - wq ) avg + wq * q1 avg = wq * q1 Second queue length measure: q2 avg = ( 1 - wq ) avg + wq * q2 avg = ( 1 - wq ) wq * q1 + wq * q2 Third queue length measure: q3 avg = ( 1 - wq ) avg + wq * q3 avg = ( 1 - wq ) [ ( 1 - wq ) wq * q1 + wq * q2 ] + wq * q3 avg = ( 1 - wq )2 wq * q1 + ( 1 - wq ) wq * q2 + wq * q3

  ---

  ---

• Relationship between current queue length and exponential weighted average queue length
  • The following graph shows the relationship between the current queue length (the darkblue plot) and the exponential weighted average queue length (the red plot):

  • The average queue length is more smoother than the current (instantaneous) queue length

  ---

  ---

  ---

  ---

• Drop Probability: linearly increasing function

  • The drop probability Prob(queue length) increases linearly from 0 ... max_p between queue length values min_threshold ... max_threshold:

    ---

  • marking (drop) algorithm:

    count = number of UNMARKED packets since the last marked packet if ( avg < minthreshold ) { admit new packet; } else if ( minthreshold < avg < maxthreshold ) { /* ------------------------------------------ Compute drop probability (linear) ------------------------------------------ */ avg - minthreshold pb = --------------------------- * maxp maxthreshold - minthreshold /* ---------------------------------------------------- Adjust drop probab. when there are no drops ---------------------------------------------------- */ count++; // # times that RED fails to drop a packet pb pa = --------------- // pa increases when count increases 1 - pb * count /* ------------------------------------------ Drop packet with probab pa ------------------------------------------ */ if ( random() < pa ) { Mark arriving packet count = 0; // Reset count... reduce adjusted drop probability pa } } else if ( avg > maxthreshold ) { Mark arriving packet count = 0; }

  ---

  ---

  ---

• RED in NS2
  • The RED queue management algorithm is implemented in NS2 as the class REDQueue : click here

    The RED drop code is found around line 508 in the file red.cc

  ---

  ---

• Evaluation of the Performance of RED queue management
  • The paper "Random Early Detection Gateways for Congestion Avoidance" presented a simulation study of the RED algorithm.

  • Simulation network used:

    • Node 1 begins to transmit at time 0.0 sec
    • Node 2 begins to transmit at time 0.2 sec
    • Node 3 begins to transmit at time 0.4 sec
    • Node 4 begins to transmit at time 0.6 sec

    ---

  • Average Queue Size in congested router:

    • The instantaneous queue size is given by the jagged line
    • The average queue size is given by the smooth line

    • The weight parameter w_q = 0.002

    • The parameters min_threshold = 5 packets and min_threshold = 15 packets
      They are given by the 2 dotted lines in the figure

    • The parameters max_p = 1/50 (0.02)

    ---

  • Results of the simulation:

    • NO synchronized packet drops among the flows.

  ---

  ---

• Further Benefits of RED
  • The following simulation study shows that the RED queue management can achive higher throughput than Drop Tail

    ---

  • The simulation scenario:

    ---

  • The throughput vs. queue length in RED and Drop Tail:

  • The authors attribute the improved throughput to the reduced average queue length

  • Packets wait less time in the queue and get out faster.

    (Analogy: keep the car on the high way moving and you can get more cars through. Longer lines not only slows the traffic, in fact, a lower number of car per time unit will get through to the destination)

  ---

  ---

• Parameter Sensitivity in the RED method
  • There are 4 important parameters in RED:
    1. min_threshold
    2. max_threshold
    3. max_p
    4. w_q

  • RED is exceptionally sensitive to the setting of its parameters

  • It is very hard to tell which values are right for a particular case... (sort of trial and error if you don't practice black magic...)

  ---

  ---

• Recommended Parameter Setting (see: http://www.icir.org/floyd/REDparameters.txt)

  • The following guides are suggested in picking the following RED parameters:

    maxthreshold = 3 * minthreshold           maxp = 0.1

  • There are no definitive recommendations on how to set min_threshold and w_q

  • The authors do suggest the following (rather strange) recommendations:

    minthreshold = 5 packets ???           wq = 0.002

  ---

  ---

• Simulating RED Queues in NS2
  • A RED queue is defined between nodes $n1 and $n2 using the RED queueing option.

    Example:

    $ns duplex-link $n1 $n2 10Mb 10ms RED

    ---

  • The RED queue has a number of parameters that you can set.

    The name of the parameters (OTcl instance variables) are:

    thresh_ = minumum threshold (queue threshold at which RED begins to mark packets probabilistically) maxthresh_ = maximum threshold (queue threshold at which RED begins to mark every arriving packets) q_weight_ = wq (factor in the Exponential Weighted Moving Average) linterm_ = 1/maxp (as average queue size varies between thresh_ and maxthresh_, the packet dropping probability is increased from 0 to 1/linterm_ setbit_ = marking/dropping flag. If setbit_ = 0 or false, a marked packet is simply marked and not dropped. If setbit_ = 1 or true, a marked packet is dropped

    Examples:

    Queue/RED set q_weight_ 0.002 Queue/RED set thresh_ 10 Queue/RED set maxthresh_ 20 Queue/RED set setbit_ true

    ---

  • DEMO: (RED vs. DropTail)                        
    • 2 Reno flows in DropTail: click here

    • 2 Reno flows with RED: click here

      Try changing the Min Threshold and Max Threshold.

    ---

  • CWND Output from these programs:

    DropTail experiment:

    ---

    RED experiment:

  ---

  ---

• Epilogue
  • RED is very controversial...

  • Publications like: "Reasons not to deploy RED" (1999) testify against the use of RED: click here

    Setting the parameters in RED is more like Black Magic than anything else...

    And this is supposed to work in a WORD WIDE Internet - not gonna happen.