WRED on the EQ DPC

Shaping to drop out-of-profile traffic is done on the EQ DPC at all levels but the queue level. However, Weighed random early discard (WRED is done at the queue level with much the same result. With WRED, the decision to drop or send the packet is made before the packet is placed in the queue.

WRED shaping on the EQ DPC is similar to the IQ2 PIC, but involves only two levels, not 64. The probabilistic drop region establishes a minimum and a maximum queue depth. Above the minimum and below the maximum queue depth, the drop probability is 0 (send). About the maximum level, the drop probability is 100 (certainty).

There are four drop profiles associated with each queue. These correspond to each of four loss priorities (low, medium-low, medium-high, and high). Sixty-four sets of four drop profiles are available (32 for ingress and 32 for egress). In addition, there are eight WRED scaling profiles in each direction.

An EQ DPC drop profile for expedited forwarding traffic might look like this:

Note that only two fill levels can be specified for the EQ DPC. You can configure the interpolate statement, but only two fill levels are used. The delay-buffer-rate statement in the traffic control profile determines the maximum queue size. This delay buffer rate is converted to a packet delay buffers, where one buffer is equal to 512 bytes. For example, at 10 Mbps, the EQ DPC will allocate 610 delay buffers when the delay buffer rate is set to 250 milliseconds. The WRED threshold values are specified in terms of absolute buffer values.

The WRED scaling factor multiples all WRED thresholds (both minimum and maximum) by the value specified. There are eight values in all: 1, 2, 4, 8, 16, 32, 64, and 128. The WRED scaling factor is chosen to best match the user configured drop profiles. This is done because the hardware supports only certain values of thresholds (all values must be a multiple of 16). So if the configured value of a threshold is 500 (for example), the multiple of 16 is 256 and the scaling factor applied is 2, making the value 512, which allows the value of 500 to be used. If the configured value of a threshold is 1500, the multiple of 16 is 752 and the scaling factor applied is 2, making the value 1504, which allows the value of 1500 to be used

Hierarchical RED is used to support the oversubscription of the delay buffers (WRED is only configured at the queue, physical interface, and PIC level). Hierarchical RED works with WRED as follows:

• If any level accepts the packet (the queue depth is less than the minimum buffer level), then this level accepts the packet.
• If any level probabilistically drops the packet, then this level drops the packet.

However, these rules might lead to the accepting of packets under loaded conditions which might otherwise have been dropped. In other words, the logical interface will accept packets if the physical interface is not congested.

Due to the limits placed on shaping thresholds used in the hierarchy, there is a granularity associated with the EQ DPCs. The shaper accuracies differ at various levels of the hierarchy, with shapers at the logical interface level (level 3) being more accurate than shapers at the interface set level (level 2) or the port level (level 1). Table Table 47 shows the accuracy of the logical interface shaper at various speeds on a 40-port Gigabit Ethernet EQ DPC.

Table 47: Shaper Accuracy of 40-port Gigabit Ethernet EQ DPC at the Logical Interface Level

Table Table 48 shows the accuracy of the interface set shaper at various speeds on a 40-port Gigabit Ethernet EQ DPC.

Table 48: Shaper Accuracy of 40-port Gigabit Ethernet EQ DPC at the Interface Set Level

Table Table 49 shows the accuracy of the port shaper at various speeds on a 40-port Gigabit Ethernet EQ DPC.

Table 49: Shaper Accuracy of 40–port Gigabit Ethernet EQ DPC at the Port Level.

For more information about configuring RED drop profiles, see Configuring RED Drop Profiles.