Specifications

White Paper

With the introduction of the WS-X6708 and WS-X6716, the placement of a packet within a queue can be determined

by the DSCP value of the IP packet directly. This suppresses the need for DSCP to CoS mapping and simplifies the

queue/threshold assignment process.

6.5 Weighted Random Early Discard and Round Robin Scheduling

The Weighted Random Early Discard (WRED) congestion management mechanism and the Round Robin

scheduling mechanisms are extremely powerful algorithms resident on the Catalyst 6500 port ASIC’s. There are

three implementations of Round Robin scheduling on the Catalyst 6500 and they include Weighted Round Robin

(WRR), Deficit Weighted Round Robin (DWRR) and Shaped Round Robin (SRR). WRED and all the Round Robin

scheduling options use the priority tag (CoS) inside an Ethernet frame to provide enhanced buffer management and

outbound scheduling. All of these are explained further later in this document.

6.5.1. WRED

WRED is a buffer management algorithm employed by the Catalyst 6500 to minimize the impact of dropping high

priority traffic at times of congestion. WRED is based on the RED (Random Early Discard) algorithm used in some of

our competitors switching products.

Before we look at RED and WRED, let’s quickly revisit TCP flow management. Flow management ensures that the

TCP sender does not overwhelm the network. The “TCP slow-start” algorithm (defined in RFC2001) is part of the

solution to address this. It dictates that when a flow starts, a single packet is sent before it waits for an

acknowledgement (ACK). When the ACK is received, the TCP endpoint will send two packets and wait for the next

ACK before sending more data. This process gradually increases the number of packets sent before each ACK is

received. This will continue until the flow reaches a transmission level (i.e. send “x” number of packets) that the

network can handle without the load incurring congestion. Should congestion occur, the slow-start algorithm will

throttle back the window size (i.e. number of packets sent before waiting for an acknowledgement). This will

normalize transmission to a set number of frames that the network can handle without dropping them.

Back to RED and WRED. RED will monitor a queue as it starts to fill up. Once a certain threshold has been

exceeded, packets will start to be dropped randomly. No regard is given to specific flows; rather random packets will

be dropped. These packets could be from high or low priority flows. Dropped packets can be part of a single flow or

multiple TCP flows. If multiple flows are impacted, as described above, this can have a considerable impact on each

flows window size.

Unlike RED, WRED is not so random when dropping frames. WRED takes into consideration the priority of the

frames (in the Cat6500 case it uses to CoS value). With WRED the administrator assigns frames with certain CoS

values to specific thresholds. Once these thresholds are exceeded, frames with CoS values that are mapped to

these thresholds are eligible to be dropped. Other frames with CoS values assigned to the higher thresholds are en-

queued. This process allows for higher priority flows to be kept intact keeping their larger window sizes intact and

minimizing the latency involved in getting the packets from the sender to the receiver.