3Com 86-0621-000, C36460T software manual Class-Based Queuing CBQ Management

16ENTERPRISE OS SOFTWARE VERSION 11.4 RELEASE NOTES

■Class-Based Queuing (CBQ) Management

Class-Based Queuing (CBQ) is a link-sharing packet scheduler which is an enhanced version of the existing Protocol Reservation queuing policy. It performs priority scheduling and supports specific traffic class characteristics, such as the average transfer rate. It supports a hierarchy of service classes, each associated with a set of QoS attributes (such as, average rate, priority, and max delay) and a packet queue to hold packets marked for the service class.

CBQ provides weighted (based on the allocated bandwidth) round robin scheduling when the class is not congested, but switches to the link sharing mode during periods of congestion. It regulates each class queue to its allocated bandwidth, but allows a congested class to borrow bandwidth from its under-utilized parent class.

When a class queue builds up due to packet arriving at higher rate than the class’s allocated bandwidth, CBQ employs a packet drop policy to manage the queue length/latency. By default, the simple “tail drop” is invoked to discard the most recently arrived packet for the congested queue/class. The more effective RED dropper can also be optionally enabled on a CBQ class queue.

CBQ also supports traffic prioritization. Higher priority classes are serviced first, classes with the same priority are then serviced based on weighted round robin. Borrowing is allowed only if a class is configured to allow borrowing from its parents.

The network manager may define any number of CBQ classes. Policies can be defined that map the DSCP in the TOS-byte to a specific service class to provide the desired QoS. Initial RSVP support will restrict RSVP flows to the well-known “RSVP” service class.

Given the significant per packet overhead, CBQ does not scale well with multi-level class hierarchies and would perform best with a small number of classes in a shallow tree structure on lower speed WAN links.

CBQ will be supported on PPP/FR ports only.

■RED Congestion Avoidance

Random Early Discard (RED) actively manages the queue size by dropping arriving packets using probability as follows. The probability of packet drops increases as the estimated average queue size grows. The average queue size is computed using a simple exponentially weighted moving average estimator. RED starts dropping arriving packets when the queue size exceeds the defined minimum threshold in number of packets), and the drop probability increases linearly with the queue size until the defined maximum threshold (in number of packets) is reached - at which point all arriving packets are dropped.

Weighted Random Early Discard (WRED) implements an additional drop-precedence based preferential discard mechanism. The drop-precedence value is used to determine the minimum and maximum thresholds–such that packets tagged with a higher drop-precedence value has a higher drop probability. The drop-precedence value is determined by the amount of traffic in excess of the rate limit.

RED congestion avoidance scheme actively manages the queue length to efficiently reduce both packet drops and queue latency, resulting in lower delay and better service. The random packet drop also effectively breaks up the traffic synchronization due to TCP’s “slow start than speed up” behavior, which