India Telecom

February 24, 2005
Deploying QoS in the enterprise

Mathew Varghese

Operating a converged enterprise network requires some optimization to guarantee that all applications receive the service levels required to meet performance expectations. There is a common misconception that purchasing an oversupply of bandwidth will solve all service-quality challenges. "Throwing bandwidth at the problem" is sometimes perceived as a simpler solution than quality-of-service (QoS) management

NEW DELHI -- Adequate bandwidth provisioning is one important aspect of QoS, as having sufficient bandwidth helps minimize congestion. However, the selfish behavior of TCP-based applications makes relying on capacity alone incomplete as a performance-management strategy. The reason being, as TCP speakers gradually, yet continually, increase their transmission rates, consuming all available bandwidth, and slow down only when they detect packet loss. This process is called TCP slow start as defined in Internet Engineering Task Force (IETF) RFC 2001. It can detract from the performance of the other applications if they are not protected by QoS.

Good QoS design begins with identifying low-latency applications and marking them for high-priority treatment throughout the network. Additional applications requiring bandwidth guarantees should be subsequently marked and protected. Finally, applications hogging bandwidth might be identified and tamed as well.

FIGURE: Voice and video applications are less tolerant of loss, delay, and delay variation (jitter) than data, but their requirements are more straightforward. Data applications vary widely in their QoS requirements, so they should be profiled before determining appropriate classification and scheduling treatment.

There are three primary configuration components to QoS:

Classification: Marking packets so that varying service levels can be enforced throughout the network;
Scheduling: Assigning packets to one of multiple queues and associated service types based on the classification for specific service level treatment by the network; and
Resource provisioning: Accurately calculating the required bandwidth for all applications plus overhead (see Figure). Let's see why these steps are necessary and where and how they should be implemented.

QoS at network intersections
QoS features should be configured in multi-service networks where there are speed mismatches between interconnected circuits and where many-to-one aggregation occurs:

• On access-switch ports connected to end stations, such as IP phones, voice gateways, and voice application servers;
• On network interfaces interconnecting access-layer and distribution-layer switches within a campus network;
• On network interfaces interconnecting distribution-layer and core switches within a campus network; and
• In WAN access routers linking multi-megabit- or gigabit-speed LAN segments to lower-speed WAN circuits, such as a T1/E1 or T3/E3, at branch offices or central aggregation sites.

At these network intersections, traffic can be throttled back to a lower speed, and link oversubscription occurs. This is where packets are most likely to be queued as they wait to transmit. As packets queue up, the potential for latency and packet loss arises. The appropriate number of queues and algorithm(s) used to service each queue differs, depending on device and network location. Queue management or scheduling lies at the heart of the QoS.

Impact of network metrics
If network routers and switches have not been provided with QoS rules, traffic might be adversely affected by network congestion in the following ways:

Packet loss: If there is no special priority as to which packets should be discarded from an overflowing queue, high-priority packets might wind up being the ones to drop off. Even the discard of two VoIP packets in a row is noticeable by a user in the form of audible clips, missing pieces of conversation, or even white noise at the beginning of speech. Although video is a bit more tolerant of loss and delay, in context of a real-time, interactive conferencing (IP/VC) session, the voice component is affected the same as VoIP.

Delay: Coding, serialization, and propagation delay are among the many contributors to network delay, a metric that also affects voice conversations. Excessive delay does not affect the quality of what the user hears; it affects the quality of the conversation by introducing the inability for a user to interrupt the party at the other end.

Delay variation (jitter): Packet voice conversations are least tolerant of jitter. To compensate, special jitter buffers built into software on some VoIP phones and gateways, collect all voice packets on one side of a conversation and forward them in an even beat to the listener. Jitter buffers require no special configuration on the part of the network manager.

Trusted and untrusted classifications
When classifying traffic types in an enterprise network, a trust boundary must be established. The boundary is established by the access device, which either classifies traffic that it allows into the network itself or trusts classification that has already been applied by an end station, such as an IP phone.

Some enterprise switches are equipped to trust class-of-service (CoS) markings in an IEEE Layer 2 802.1Q/p environment, as well as those implemented at Layer 3 using Differentiated Service Code Point (DSCP) or IP Precedence. Layer 3 options are used when more granularity in marking is required than is available at Layer 2 or when the Layer 2 media changes between end stations (for instance, when traversing a WAN). Differentiated Services (DiffServ) framework and its associated 6-bit DSCP field were created as successors to IP Precedence, which uses only three bits in the type-of-service (ToS) byte in an IP header to prioritize traffic. IP Precedence enables the creation of eight service classes, compared with the 64 classes possible in the DiffServ model.

Within DSCP, the IETF has specified three broad classes of per-hop behaviors (from node to node) to ensure service-level consistency among different manufacturers' equipment: Expedited Forwarding (EF) for delay-sensitive priority queuing, Assured Forwarding (AF) for intermediate levels of preferred servicing, and Best Effort (BE). Use of DSCP is recommended wherever possible, so that devices do not require configuration changes if a greater degree of classification granularity is eventually needed.

Designing the campus
The campus network comprises two or more "hierarchical layers" of LAN switching devices, depending on the size and design chosen by the enterprise.

The access layer and IP telephony settings: A base set of features in access-layer switches is required to support IP telephony alongside data traffic in the LAN. These features include the ability of the switch to support multiple VLANs on the access port connected to an IP phone (so that a 'voice VLAN' can be set up) and the ability to manipulate the IP phone's trust boundary and marking capabilities.
Distribution and core layers: Switches are commonly deployed in the distribution and core layers of a campus network. In this area of the network, port level keywords trust dscp or trust cos are used to allow classification applied at the edge of the network in the access layer to continue throughout the network.

Configuring WAN access router
For taming bandwidth-hog applications, queuing with aggressive rate limiting during periods of network congestion is preferable to policing in the WAN. However, first these bandwidth-hogging applications must be identified, which can be done by Network-Based Application Recognition (NBAR) classification engine in routers. NBAR can also determine how much bandwidth each type of traffic is consuming to expose the 'top-talking' applications. Aggressive dropping of bandwidth hogs is preferred to policing, as policing discards packets that exceed the budget for a particular application or class. Queuing techniques such as Low Latency Queuing (LLQ) and Class-Based Weighted Fair Queuing (CBWFQ) allow bandwidth-hungry applications to use as much bandwidth as they want when there is no contention for the network. This is helpful for applications such as large file transfers or backup operations that typically occur at night.

To sum up QoS tools of classification, scheduling, and provisioning are used to minimize the impact that loss, delay, and delay variation have on real-time applications such as VoIP and videoconferencing. In addition to providing such protection, QoS tools can rate limit bandwidth-hogging applications during periods of congestion and guarantee mission-critical applications the services they require -- making a true integrated voice, video, and data network possible.