El género Populus

Until recently, the conventional wisdom was that QoS would never be an issue in the enterprise campus due to the bursty nature of network traffic and the capability of buffer overflow. However, we now know that buffering, not bandwidth, is the primary issue in the campus.Therefore, QoS tools are required to manage these buffers to minimize loss, delay, and delay variation.

Figure 4-11 QoS Problems Areas (1 of 3)

Transmit buffers have a tendency to fill to capacity in high-speed campus networks due to the bursty nature of data networks combining with the high volume of smaller TCP packets. If an output buffer fills, ingress interfaces are not able to place new flow traffic into the output buffer. Once the ingress buffer fills (which can happen very quickly), packet drops will occur. These drops will typically be more than a single packet in any given flow. Current Cisco DSP algorithms can correct for 30 Msec of lost voice. Cisco VoIP technology uses 20 Msec samples of voice payload per VoIP packet, which means that only a single voice RTP packet can be lost during any given time period. If two successive voice packets are lost, voice quality will degrade.

Figure 4-12 QoS Problem Areas (2 of 3)

VoIP traffic is sensitive to delay and drop. As long as a campus uses Gigabit Ethernet trunks, which have extremely fast serialization times, delay should never be a factor, regardless of the size of the queue buffer. However, drops will adversely affect voice quality in the campus. Using multiple queues on

IP 48172 Si Si Si Distribution Access Potential QoS problem areas Core TX TX TX TX TX TX TX TX TX Data Data To core

Ethernet switch Data flows "hog" Tx buffer

Additional flows, including voice, can not get access to Tx queue Data Voice RX RX TX RX RX 48152

transmit interfaces is the only way to eliminate the potential of drops caused by buffers operating at 100 percent capacity. By separating voice and video into their own queues, flows are never dropped at the ingress interface if data flows fill up the data transmit buffer.

Figure 4-13 QoS Problem Areas (3 of 3)

Note It is critical to remember to verify Flow Control is disabled when enabling QoS (multiple queues) on Catalyst switches. Flow Control will interfere with the configured queuing behavior by acting on the ports before queuing is activated. Flow Control is disabled by default.

The scheduler process can use a variety of methods to service each of these transmit queues. The easiest method is a Round-Robin (RR) algorithm, which services queue 1 through queue N in a sequential manner. While not robust, this is an extremely simple and efficient method that can be used for branch office and wiring closet switches. Distribution Layer switches use a Weighted Round-Robin (WRR) algorithm so higher priority traffic is given a scheduling weight. Another option is to combine RR or WRR scheduling with priority scheduling for delay and drop sensitive applications. This uses a priority queue (PQ), which, when there are packets in the queue, is always serviced first. If there are no frames in the PQ, then the additional queues are scheduled using RR or WRR.

An important consideration is how many queues are actually needed on transmit interfaces in the campus:

• Should you add a queue to wiring closet switches for each CoS value? • Should you add eight queues to the Distribution Layer switches? • Should you add a queue for each of the 64 DSCP values?

It is important to remember that each port has a finite amount of buffer memory. A single queue will have access to all the memory addresses in the buffer. As soon as you add a second queue, the finite buffer amount is split into two portions, one for each queue. All frames entering the switch not classified for entry into the newly created second queue are now contending for a much smaller portion of buffered memory registers. Therefore, during periods of high traffic, the buffer will fill and frames will be dropped at the ingress interface. Considering that the vast majority of network traffic is TCP-based (comprising TCP ACKs (40 Bytes), TCP SYN/ACKs (44 Bytes) and 512-1024 Byte TCP application traffic (SMTP, HTTP, FTP)), dropping a packet results in a re-send. In other words, dropped packets within TCP-oriented networks increase network congestion. Queuing should be used cautiously and only when particular drop and delay sensitive priority traffic is traversing the network.

Data Data

To core

Ethernet switch Queue assignment based on CoS/ToS classification

Voice put into "delay and drop" sensitive queue Data Voice RX RX TX RX RX 48153

Two queues are adequate for wiring closet switches, where buffer management is less critical. Whether these queues are serviced in a RR, WRR, or Priority Queuing manner is less critical because the scheduler process is extremely fast when compared to the aggregate amount of traffic.

Distribution switches require much more complex buffer management because of the flow aggregation occurring at this layer. Not only do you need priority queues, but you also need thresholds within the standard queues.

Cisco chose to use multiple thresholds within queues instead of continually increasing the number of interface queues. Each time a queue is configured and allocated, all of the memory buffers associated with that queue can only be used by frames meeting the queue entrance criteria. For example, we will assume that a Catalyst 4000 10/100 Ethernet port has two queues configured: one for VoIP (VoIP bearer and control traffic) and the default queue which is used for HTTP, e-mail, FTP, logins, NT Shares, and NFS. The 128KB queue is split into a 7:1 transmit and receive ratio. The TX buffer memory is then further separated into high and low priority partitions in a 4:1 ratio. If the default traffic (the web, E-mail, and file shares) begins to congest the default queue, which is only 24KB, then packets begin dropping at the ingress interfaces regardless of whether or not the VoIP control traffic is using any of its queue buffers. The dropped packets of the TCP-oriented applications will cause each of these applications to re-send the data again, aggravating the congested condition of the network. If this same scenario were configured with a single queue, but multiple thresholds used for congestion avoidance, then the default traffic would share the entire buffer space with the VoIP control traffic. Only during periods of congestion, when the entire buffer memory approaches saturation, would the lower priority traffic (HTTP and e-mail) be dropped.

This is not to say that multiple queues are not a vital component in IP Telephony networks. As discussed earlier, the VoIP bearer streams must use a separate queue to eliminate the adverse affects of drops and delays to voice quality. However, every single CoS or DSCP value should not get its own queue because the small size of the resulting default queue will cause many TCP re-sends and actually increase congestion.

The VoIP bearer channel is also a poor candidate for queue congestion avoidance algorithms like Weighted Random Early Detection (WRED). Queue thresholding uses the WRED algorithm to manage queue congestion when a preset threshold value is set. WRED works by monitoring buffer congestion and discarding TCP packets if the congestion begins to increase. The result of the drop is that the sending endpoint detects the dropped traffic and slows the TCP sending rate by adjusting the window size. A WRED drop threshold is the percentage of buffer utilization at which traffic with a specified CoS value is dropped, leaving the buffer available for traffic with higher-priority CoS values. The key is the word “Random” in the algorithm name. Even with Weighting configured, WRED can still discard packets in any flow; it is just statistically more likely to drop from the lower CoS thresholds.