QoS Requirements

The service requirements or the QoS guarantee provided by the network are the terms of the network's performance. Bandwidth, packet delay, jitter, and packet loss are some common measures used to characterize a network's performance. The QoS requirements vary depending on the applications' requirements. For VoIP or IP Telephony, packet delay, jitter, and packet loss are important. For applications that involve bulk data transfer, bandwidth is a QoS requirement. Additionally, for bulk data, the requirement is to constrain bandwidth (that is, not to permit a few TCP sessions to adversely dominate a link to the detriment of interactive applications).

Bandwidth

Bandwidth describes the throughput of a given medium, protocol, or connection. It describes the size of the pipe that is required for the application to communicate over the network. An application requiring guaranteed bandwidth wants the network to allocate a minimum bandwidth specifically for it on all the links through which the application's data is transferred through the network. Depending on the type of network, the bandwidth guarantee can be provided at the IP layer or the data link layer. Guaranteed bandwidth at the IP layer depends on the type of data-link network. Not all data-link networks support guaranteed bandwidth when several IP connections share the same data link network.

Packet Delay and Jitter

Packet delay or latency at each hop consists of the following:

• Serialization or transmission delay How long it takes a device to send the packet at the output rate. This depends on the packet's size and the link bandwidth. A 64-byte packet on 4-Mbps line takes 128 μs to be transmitted. The same 64-byte packet on a 128-kbps line takes 4 ms to be transmitted.

• Propagation delay How long it takes a bit to be transmitted by the transmitter and to be received by the receiver. This is a function of the media and the distance and is independent of the bandwidth.

• Switching delay How long it takes a device to start transmitting a packet after it receives the packet. This depends on the network's status and the number of packets in transit at this hop.

End-to-end delay for a packet belonging to a flow is the sum of all these types of delays experienced at each hop. Not all packets in a flow will experience the same delay. It depends on the transient delay in each hop in the network. If the network is congested, queues are built at each hop, and this increases the end-to-end delay. This variation is in the delay and is called jitter.

Queuing mechanisms at each node can be used to ensure that the delay of certain flows is minimized and that the delay jitter has an upper bound, which is a defined binding of a degree of variation, as in delay and jitter.

Packet Loss

Packet loss specifies how many packets are lost in the network during transmission. Packet loss can be caused by corruption in the transmission medium, or packets can be dropped at congestion points in network due to lack of buffer space in the incoming or outgoing interface. Packet loss caused by drops should be rare for a well-designed network that is correctly subscribed or undersubscribed. Packet loss caused by faulty transmission media can be avoided by building good physical networks.

Note

Packet loss during noncongestion is a function of network availability, such as designing a network to be highly available. (Five-nines is a target, meaning that the network has less than 5 minutes downtime/drops per year.) The loss due to congestion is inherent with networks that are based on oversubscription and speed mismatches. QoS design scope addresses loss caused by congestion via congestion-avoidance mechanisms.

Enterprise Loss, Latency, and Jitter Requirements

VoIP and video applications are sensitive to delay and jitter. Delay is the amount of time taken by the IP network to deliver the packet from the source to the destination. Jitter is the variation in the delay. Unlike traditional IP-based applications that depended on best-effort services, VoIP applications have strict delay and jitter requirements. Packets from these applications must be delivered to the destination with a finite delay (about 150 ms). Videoconferencing has a built-in VoIP component (typically G.711) and therefore has the same delay and jitter requirements as traditional VoIP. Note that the 150-ms end-to-end target is an ITU-T specification, G.114. VoIP and videoconferencing require guaranteed end-to-end bandwidth, meaning that at any time, the IP network can guarantee a minimum throughput (measured in kbps) from the source to the destination. In the world of data communication, the need for communication is to access data or to exchange data. Data may be located in a central location or may be located in several locations. Depending on the nature of the data and the type of applications that require the data, the QoS requirements of the data communication can vary.

Traditionally, IP networks have provided best-effort service. Best-effort means that the packet may eventually be delivered to the destination. There is no guarantee that the packet will be delivered within a given time period or even be delivered at all. However, there is an implied "good-faith" guarantee that best-effort traffic will be delivered. IP provided the means for delivering data from a source to one or multiple destinations. The Internet is a classic example of this type of IP network. The transport layer (TCP) provided QoS such as the following:

• Guaranteed delivery in case of packet loss (retransmission)

• Delivery of packets in the correct order (reordering of packets)

• End-to-end flow control in case of congestion in the end systems or in the service provider network

In addition to best-effort, some of the additional requirements are as follows:

• Delivery on time (low latency)

• Delivery of packets with minimum loss (low loss)

• Delivery of a minimum number of packets in a given time period (guaranteed bandwidth)

• Data security

The first two requirements can be fulfilled using the transport layer (TCP), but the other requirements depend on IP and the lower layers used by IP to deliver packets. Delay is how long it takes the packet to reach the destination. Delay variation refers to the difference in delay for each packet that is delivered to the same destination. Packet loss means that not all packets are delivered to the destination.

Delay depends on the queuing mechanisms IP uses to deliver packets when there is congestion on the network. It is possible to compute the minimum delay to deliver a packet from the source and destination based on the information about the physical media and Layer 2 services and the delays in the intermediate nodes in the network. However, congestion in the network can result in additional delays in the intermediate nodes and can increase the end-to-end delay.

Note

End-to-end delay is how long it takes the packet to be delivered from the source to the destination. It also includes the time taken to retransmit the packet if the packet gets lost in an intermediate node. Several factors affect delay, such as fixed packetization, fixed propagation, variable serialization, and variable queuing.

Some of the reasons for packet loss are faults that occur at Layer 2, faults that occur at Layer 1, and network congestion. Network congestion can occur because of lack of resources such as memory to buffer incoming packets, and also when the sum of bandwidth for all the incoming interfaces exceeds the bandwidth of the outgoing interface. IP depends on the QoS offered by the lower layers in providing this QoS to the application layer services. Overall, you can expect network congestion due to natural oversubscription that is integrated into networks, such as speed mismatches.

QoS at Layer 2

Depending on the QoS requirements, QoS functions are available at the data link layer (Layer 2) and network layer (Layer 3) of the Open System Interconnection (OSI) reference model. Guaranteed bandwidth as a QoS requirement can be provided by several Layer 2 technologies, such as Frame Relay and ATM, when the physical medium is shared simultaneously by several Layer 3 connections. ATM can also meet other QoS requirements, such as delay and jitter. Furthermore, guaranteed bandwidth can be provisioned on Layer 2 protocols via InterServ QoS, such as Resource Reservation Protocol (RSVP). DiffServ does not provide explicit bandwidth guarantees, but rather per-hop behavior constructs. Layer 2 protocols in themselves do not provide bandwidth guarantees. For example, it is possible to oversubscribe both Frame Relay and ATM and cause drops.

QoS is an important aspect of any IP service offering. QoS helps define an important component of the SLApacket delivery. When it comes to delivery of business-critical traffic, in most cases, best-effort delivery is no longer sufficient. Many applications require bandwidth guarantees, and some of them also require delay and jitter guarantees. Any service delivered with Layer 3 must also be able to deliver QoS to meet the needs of business-critical applications. Applications such as voice and video have stringent requirements with respect to jitter and packet loss. These days, highly efficient codecs can mask some delay and jitter with appropriate buffering, encoding, and decoding techniques. Nevertheless, despite these efficient codes, bandwidth, delay, and jitter guarantees are still needed from the network for better quality of experience (QoE). Because a large number of enterprises and services are considering IP/MPLS for network convergence, the expectations for an IP/MPLS network are very high. Many times, we have seen a comparison between a Frame Relay or ATM network and Frame Relay or ATM QoS. Enterprises are used to the bandwidth models of Frame Relay committed information rate (CIR) and ATM constant bit rate (CBR) (also called guaranteed bit rate service). They use these commonly as access circuits into the provider network, even for IP access. They have an expectation of peak and sustained information rate based on the connection-oriented nature of Frame Relay or the ATM network.

The QoS delivered by the subscriber network results from the network design. In migrating to an MPLS VPN, problems in the design of the subscriber network are more likely to affect QoS than issues in the service provider backbone.

SLAs are contracts with the service provider guaranteeing the level of service the VPN core will deliver. Effective SLAs include quantifiable measurements of service levels and remediation requirements.

The subscriber network design and the backbone SLA are tightly coupled in a subscriber MPLS VPN. In legacy networks based on dedicated circuits or permanent virtual circuits (PVCs), QoS and SLAs by definition were designed and enforced on a site-to-site basis. In an MPLS VPN, the link from each site to the core contracts QoS individually. Mismatches in the traffic load may cause unacceptable performance in the subscriber network, while the core is meeting all the QoS requirements.

The full mesh of PVCs, and sometimes multiple PVCs for differing levels of QoS in the existing subscriber network, is replaced with a single link at each site. With bandwidth dynamically shared across all classes of service (CoSs), efficiency and economy are optimized in the subscriber network. Frequently, MPLS VPNs perform so well that attention to network design gradually diminishes. Migrating or adding new sites is reduced to ordering a new link with the same characteristics as existing sites.

Designing and maintaining the QoS in a subscriber MPLS VPN requires understanding the volume, characteristics, and patterns of each of the traffic load classes, as well as vigilance in expanding the network for additional applications.

IP QoS implementation can be divided into the following categories:

• Classification The use of Differentiated Services Code Point (DSCP) or the first 6 bits of the Type of Service (ToS) field classifies traffic based on customer requirements. As soon as the packets are properly classified, they can be accordingly handled by other QoS mechanisms, such as congestion management and policing and congestion avoidance to properly implement end-to-end QoS requirements. Packet classification typically is done on the edge of the network. Sometimes, packets are reclassified in the core network by remarking certain fields in the packet. This reclassification is required when traffic is aggregated. However, the network must ensure that the original value of the DSCP field in the IP packet is restored at the edge of the network when the packet is delivered to the customer. This can be done in an MPLS network because two fields are available.

• Congestion management This involves creating queues, assigning packets to the proper queues, and scheduling the queues and the packets within the queues. The number of queues depends on the customer requirements and the number of CoSs offered by the network. Assigning packets to queues and scheduling policies depend on the type of QoS services offered to the customer. For example, high-priority traffic such as VoIP requires preemptive queue mechanisms that ensure that VoIP packets are scheduled and transmitted before other packets. RFC 3246 describes this mechanism in detail.

• Congestion avoidance techniques Congestion avoidance is achieved by dropping the packets. Which packets have to be dropped are determined based on the drop policy. Congestion avoidance mechanisms are activated when congestion is detected (that is, to the point when congestion buffers reach capacity). Congestion avoidance techniques do not ensure that the network is congestion-free; rather, they obviate TCP slow-starts and TCP synchronization by randomly dropping packets when queuing buffers fill. Each queue has Weighted Random Early Detection (WRED) thresholds. Drops commence when these thresholds are exceeded. If substantial traffic is offered to a queue with a guaranteed bandwidth via class-based weighted fair queuing (CBWFQ), for example, it could trigger WRED drops before the WRED threshold is exceeded on the best-effort queue. Each WRED mechanism is queue-dependent and therefore is independent of other queues.

• Policing and shaping mechanisms These mechanisms ensure that each CoS (based on the marked IP packet) adheres to the service contract. The service contract can include several issues, such as bandwidth, burst size, and delay. A policer typically drops or remarks out-of-contract traffic, and a shaper delays excess traffic.

• QoS signaling This is used between nodes in the network to signal the QoS requirements of each class and also to reserve resources. RSVP is a QoS signaling protocol that can be used to reserve resources such as bandwidth. QoS signaling mechanisms also depend on the routing protocols to determine the best path between the source and the destination.

Implementing QoS in an IP network is a challenging task. It requires a good understanding of queuing theory and the customers' requirements to determine the parameters for the queuing policies. Some of the challenges are communication between the signaling plane (QoS signaling protocols, such as RSVP) and the data-forwarding plane (congestion in network) to ensure that resource reservation for an application can be done correctly. For example, RSVP uses bandwidth as the resource to do reservation. In addition to bandwidth, other network resources, such as queue buffers on the network devices, are also important resources that are required to guarantee QoS. Congestion in the network device due to lack of queue buffers must be communicated to RSVP so that it can use alternative paths (between the source and destination) that have enough network resources (bandwidth, queue buffers, and so on) to meet the QoS requirements of the application making the RSVP request.