Paper review: Analysis and Design of an Adaptive Virtual Queue (AVQ) Algorithm for Active Queue Management

Reviewer: Kevin Hofstra

1. What is the best Active Queue Management method? Should indications to the end nodes of congestion be in the form of marking the packets or by just dropping them? What should be considered optimal between goodput and delay latency?
2. A comparative analysis of several distinct Queue management techniques. An introduction to an Adaptive Virtual Queue method that uses packet marking for notifications and a virtual queue for decisions.
3. A. The different queue management methods (Type of notification):

i. AVQ(Packet marking)

ii. GKVQ (Packet marking)

iii. RED (Packet dropping)

iv. PI (Packet dropping)

2. AVQ uses Explicit Congestion Notification to notify users of congestion. When it detects the onset of congestion it marks a bit in the packet header that the end node interprets as a dropped packet, without having to actually drop the packet. This increases the goodput by reducing retransmissions.
3. The algorithm maintains a virtual queue that has a smaller virtual capacity than the actual buffer. Then when this queue has overrun, it is interpreted as congestion approaching the actual queue threshold, and packets are marked explicitly to slow the sending from the end users. This prevents the actual queue from ever dropping packets, but also slows users sending rates in the event of approaching congestion. The virtual queue is dynamic and does not have to depend on the size of the actual queue.

4. Critique the main contribution
• Significance- 2 The article is more of an implementation change than an actual discovery. The packet marking has already been done by GKVQ, and the active queue can for the most part be replaced by placing a threshold on a fraction of the whole actual queue. The thought of having a preliminary threshold and marking indication bits before actually dropping is far from revolutionary.
• Convincing- 2 Each of the implementations are explained and they do present some actual network data, but it is very limited in scope and not very dynamic. The conditions seem to be optimized in an effort to crowd out the good results of GKVQ and to display the weaknesses of RED and PI.
5. System researchers and builders should recognize that often the best methods are created by combining two marginal methods of creating the same results into one method. Often the extra overhead of combining them is less than the performance gain of each method.