This paper describes a technique to manage the queues at gateway level such that the queue size on average will be small but also accommodates occasional bursty traffic. This is also designed to mitigate the issue where all hosts decides to back off transmission in synchronization, as a reaction to dropped packet. This is in contrast to the traditional "tail drop" policy used at the time when the paper was first published, and is not biased towards bursty traffic.
This paper's work precedes previously discussed fair queueing work choronologically, so it was a little strange to read it in this order. This paper tries to do active queue management without keeping per flow state and differentiating between different flows. This simplifies the design at the gateway level. A good lesson from this paper is that by using active dropping packet before absolutely necessary, the gateway operates in a more optimal range and delivers packets with lower latency. Reactive systems requires the system to operate in a suboptimal state for a while before recovering, and thus are often not the best solution.
Subscribe to:
Post Comments (Atom)
1 comment:
Interesting question as to why this order. Well FQ is widely enabled in the Internet, but not RED. But that isn't really the reason. I think it is easier to understand FQ and the issues of fair bandwidth share before getting into the congestion management issue.
Post a Comment