Evaluation of Queue Management Algorithms

1
Evaluation of Queue Management Algorithms

• Ningning Hu, Liu Ren, Jichuan Chang
• 30 April 2001

2
Motivation

• Typical queuing discipline DropTail
• Active queue management
• RED (93), BLUE (99),
• FRED (97), SFB (99), CHOKe (2000)
• Problem How to compare them?
• Metrics Utilization, Fairness and Complexity
• Method simulation under common settings

3
Algorithms

• Drop Tail drop when the queue is full
• RED queue length, minth, maxth, avg
• FRED per-active-flow accounting
• BLUE loss rate, link utilization, freeze_time
• SFB identify flows using N L bins
• CHOKe compare incoming packet with random
  selected packet in queue

4
Simulation Scenario
source
destination
2Mbps
2Mbps
UDP
UDP
1Mbps, 40ms
router
router
TCP
TCP

• Topology Dumb-bell
• Metrics throughput and queue size

5
TCP
Drop Tail
RED
BLUE
TCP
FRED
SFB
CHOKe

1. TCP Flow UDP Flow 10 1
2. UDP sending rate 2Mbps
3. Buffer size 150 packets

6
Performance Comparison
Figure 1. UDP Thpt (Mbps)
Figure 2. UDP Queue Size (packet)
7
Conclusion
Algorithm Link Utilization Fairness Ease of Configuration
RED Good Unfair Hard
BLUE Good Unfair Hard
FRED Good Fair Easy (adaptive)
SFB Good Fair Easy (non-adaptive)
CHOKe Good Fair Easy (adaptive)
Algorithm Buffer size requirement Per-flow State Information Computational Overhead
RED Large No Arrival
BLUE Small No Freeze-time
FRED Small Yes Arrival-departure
SFB Large No Freeze-time
CHOKe Small No Arrival
8
FRED

• RED Per-active-flow accounting
• protection of fragile flow
• penalizing non-responsive flow
• Flow type differentiation
• Key parameter minq
• 2 or 4
• Weak point
• Compute avgq per packet

9
FRED performance
10
FRED characteristics

• Easy to configure, adaptive
• Buffer size requirement
• Will degrade as DropTail with too many flows
• Two-packet-buffer to keep fairness

11
CHOKe
CHOose and Keep for responsive flows CHOose and
Kill for unresponsive flows
12
CHOKe
13
Parameters in CHOKe
14
BLUE

• Main Point Uses packet loss and link idle events
  instead of average queue length
• Based on observation that RED is often forced
  into drop-tail mode
• Adapt to how bursty and persistent congestion is
  by looking at loss/idle events

15
BLUE Algorithm

• Pm is Blues packet dropping probability
• Upon packet loss, if no update of pm in
  freeze_time then increase by d1
• Upon link idle, if no update of pm in freeze_time
  then decrease pm by d2
• d1 gtgt d2 ?
• More critical to react quickly to increase in load

16
BLUE Characteristics

• More stable queue size and much better behavior
  with small buffers
• Good property damaged by non-responsive flows.

17
Stochastic Fair Blue(SFB)

• Objective
• Identify and penalize misbehaving flows
• Main Algorithm
• Create L hashes with N bins each
• Each bin keeps track of separate marking rate
  (pm)
• Rate is updated using standard technique(as BLUE)
  
• Flow uses minimum pm of all L bins it belongs to
• Non-misbehaving flows hopefully belong to at
  least one bin without a bad flow

18
SFB Characteristics

• Could effectively detect non-responsive flows and
  rate-limit them
• But not adaptive, the bandwidth share of
  non-responsive flows is controlled by a
  parameter(Boxtime) offered by the user.
• Boxtime is the time interval no packets from
  non-responsive flows can be enqueued.

19
SFB Unfairness of Non-responsive Flows
Our solution Randomized Boxtime a little bit.
20
Rate-limit Non-responsive flows
UDP packets
Boxtime
Queue
Time