In a traditional manner, a network wouldn't know it is congested until a packet had been dropped, and this could be a problem because the response time would have been greatly increased. The RED algorithm tries to slow down the traffic before it reaches the maximum capacity (cliff) while maintaining a high throughput.
The most effective detecting of congestion lies in the gateway itself, according to this paper, because it can distinguish between transmission delay and queueing delay. RED gateways monitor the average queue size, and either mark the congestion bit or drop a packet to signal sender's connection, depending on whether the packets are marked "essential" or "optional".
RED calculates the average queue size in the gateway by a low-pass filter and an exponential weighted moving average, and compared it to a min and max threshold. If the size is greater than max, all packets going through are marked/dropped; less than min, none. If the value falls between min and max, the probability of marking/dropping a packet is a function of average queue size.
Comment: It makes perfect sense that they drop/tag a packet based on the connection bandwidth share of a specific host. But how do we define a "host"? Is it one TCP connection, one software, one port, or one network interface card?
No comments:
Post a Comment