1 Minseok Kwon and Sonia Fahmy Department of Computer Sciences Purdue University {kwonm, fahmy}@cs.purdue.edu All our slides and papers are available at: http://www.cs.purdue.edu/~fahmy/ A Comparison of A Comparison of Load-based and Queue-based Load-based and Queue-based Active Queue Management Active Queue Management Algorithms Algorithms
19
Embed
1 Minseok Kwon and Sonia Fahmy Department of Computer Sciences Purdue University {kwonm, fahmy}@cs.purdue.edufahmy}@cs.purdue.edu All our slides and papers.
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
1
Minseok Kwon and Sonia Fahmy
Department of Computer SciencesPurdue University
{kwonm, fahmy}@cs.purdue.eduAll our slides and papers are available at:
http://www.cs.purdue.edu/~fahmy/
A Comparison of A Comparison of Load-based and Queue-based Load-based and Queue-based
Active Queue Management Active Queue Management AlgorithmsAlgorithms
• Controls average queue size.• Absorbs bursts without dropping
packets.• Prevents bias against bursty
connections.• Avoids global synchronization of TCP.• Reduces the number of timeouts in TCP.• Punishes misbehaving flows.
5
Queue-based Active Queue Management• RED [Floyd and Jacobson, 1993]
• Drops packets probabilistically in proportion to a long term average queue length (buffer occupancy).
• SRED [Ott et al., 1999]• The packet loss probability is proportional to the
instantaneous buffer occupancy and the estimated number of active flows.
• FRED [Lin and Morris, 1997]• Imposes on each flow a loss probability proportional to the
flow average and instantaneous buffer occupancy.
• BLUE [Feng et al., 2001]• Increments the packet drop probability when packet loss
occurs and decrements the packet drop probability if the link is idle.
6
Load-based Active Queue Management
• REM [Athuraliya et al., 1999]• Congestion measure (price) is computed proportionally to the
difference between input rate and output rate and current buffer occupancy at router. Source rate is computed inversely proportional to the congestion measure. Thus, the source reaches a globally optimal equilibrium.
• AVQ [Kunniyur and Srikant, 2001]• Maintains a virtual queue. When the virtual queue overflows,
packets in a real queue are marked/dropped. The virtual capacity is modified such that total flows achieve a desired utilization of the link.
• PI controller [Hollot et al., 2001]• Queue length slope determines packet drop probability and the
queue is regulated to the desired queue length.
7
Queue-based AQM: RED
Mark with PLinearly increasingFrom 0 to Pmax
No droppingor marking
Drop with P=1
Thmin ThmaxQavg
Pmax
0
Pdrop/mark
1
Average Queue Length Drop Probability P
8
Drawbacks of Queue-based AQM
• Insensitive to current queue arrival and drain rates.• Long term queue length average produces slow
response.• Difficult to configure parameters.
Small drop probability From small queue
After some period
After some period
Large queue
Small queueLarge drop probability From large queue
9
Objectives
• Is queue length a sufficient congestion indicator?
• Can we use load information for more precise congestion indication?
• Is queue length information still important, even when the load information is used?
• How do RED, SRED, FRED, BLUE, and REM compare in terms of user-perceivable metrics, such as Web response time?
• Can we achieve both high responsiveness and high throughput?
10
Load/Delay Control (LDC)
• Load factor• R = queue arrival rate / queue service rate
• Provides better load (input/output rate) response and more intuitive parameters.
• Maintains RED benefits: misbehaving flow punishment and global synchronization avoidance.
• Unlike many load-based schemes, ECN is not required.
11
Load/Delay Control (LDC)
• Uses both queue length information and load factor as multiple time scale congestion indicators.• Long-term: Queue length gives a more