Weighted Fair Queueing

The router then assigns sequence numbers to the queued packets. In the FQ algorithm, this process of sequencing the packets is optimized so that each flow gets a roughly equal share of the available bandwidth. As it receives each packet, the router assigns a sequence number based on the length of this packet, and the total number of bytes associated with this same flow that are already in the queue.

This has a similar effect to a flow-based Round Robin (RR) queueing algorithm, in which all of the flows are assigned to different queues. These queues are then processed a certain number of bytes at a time until enough bytes have accumulated for a given queue to send a whole packet. Although this is a useful way of picturing the algorithm mentally, it is important to remember that the Cisco implementations of FQ and WFQ do not actually work this way. They keep all of the packets in a single queue. So, if there is a serious congestion problem, you will still get global tail drops.

This distinction is largely irrelevant for FQ, but for WFQ it's quite important. WFQ introduces another factor besides flow and packet size into the sequence numbers. The new factor is the weight, (W), which is calculated from the IP Precedence (P) value as follows. For IOS levels after 12.0(5)T, the formula is:

W = 32768/(P+1)

For all earlier IOS levels, the weight is lower by a factor of 4,096:

W = 4096/(P+1)

Cisco increased the value to allow for finer control over weighting granularity.

The weight number for each packet is multiplied by the length of the packet when calculating sequence numbers. The result is that the router gives flows with higher IP Precedence values a larger share of the bandwidth than those with lower precedence. In fact, it is easy to calculate the relative scaling of the bandwidth shares of flows with different precedence values.