Quality of Service for Voice over IP
Available Languages
Table Of Contents
Quality of Service for Voice over IP
Packet Classification Overview
Voice Dial Peers Classification and Marking Example
Committed Access Rate Classification and Marking Example
Policy-Based Routing Classification and Marking Example
Modular QoS Command-Line Interface Classification and Marking Example
Fragmentation and Interleaving
Fragmentation and Interleaving Overview
MLP Link Fragmentation and Interleaving Example
FRF.12 Fragmentation and Interleaving Example
Frame Relay Traffic Shaping Example
Differentiated Services for VoIP
DS and the DSCP (RFC 2474, RFC 2475) Overview
Implementing DS for VoIP: Expedited Forwarding PHB (RFC 2598)
DSCP Class-Based Marking Configuration Example
Configuring Local Gateway Resources if CAC Fails
Deploying RSVP Support for LLQ
VoIP QoS over Leased Lines (Using PPP)
VoIP over Leased Line (Using PPP) Scenario
VoIP over Leased Line (Using PPP) Recommended Solution
VoIP QoS over Frame Relay Networks
VoIP QoS over Frame Relay Scenario
VoIP QoS over Frame Relay Recommended Solution
VoIP QoS over ATM Solution Using Separate Data and Voice ATM PVCs
VoIP QoS over ATM Solution Using Shared Data and Voice ATM PVCs
Quality of Service for Voice over IP
Version History
1
4/16/2001
This document was created.
2
5/15/2001
Incoporated editorial comments.
3
6/30/2001
Incorporated additional editorial comments.
Quality of Service for Voice over IP discusses various quality of service (QoS) concepts and features that are applicable to voice—in particular, Voice over IP (VoIP). This document also provides high-level examples showing how to deploy these features in different network environments.
Quality of Service for Voice over IP contains the following sections:
•
Fragmentation and Interleaving
•
•
•
•
QoS for VoIP Overview
For VoIP to be a realistic replacement for standard public switched telephone network (PSTN) telephony services, customers need to receive the same quality of voice transmission they receive with basic telephone services—meaning consistently high-quality voice transmissions. Like other real-time applications, VoIP is extremely bandwidth- and delay-sensitive. For VoIP transmissions to be intelligible to the receiver, voice packets should not be dropped, excessively delayed, or suffer varying delay (otherwise known as jitter). For example, the following standards must be met:
•
•
•
VoIP can guarantee high-quality voice transmission only if the voice packets, for both the signaling and audio channel, are given priority over other kinds of network traffic. For VoIP to be deployed so that users receive an acceptable level of voice quality, VoIP traffic must be guaranteed certain compensating bandwidth, latency, and jitter requirements. QoS ensures that VoIP voice packets receive the preferential treatment they require. In general, QoS provides better (and more predictable) network service by providing the following features:
•
•
•
•
•
Quality of Service for Voice over IP discusses various QoS concepts and features that are applicable to VoIP.
Sufficient Bandwidth
Before you consider applying any of the QoS features discussed in this document, you must first provision sufficient network bandwidth to support real-time voice traffic. For example, an 80-kbps G.711 VoIP call (64 kbps payload plus 16-kbps header) will be poor over a 64-kbps link because at least 16 kbps of the packets (which is 20 percent) will be dropped. This example also assumes that no other traffic is flowing over the link. After you provision sufficient bandwidth for voice traffic, you can take further steps to guarantee that voice packets have a certain percentage of the total bandwidth and get priority.
Packet Classification
The basis for providing any QoS lies in the ability of a network device to identify and group specific packets. This identification process is called packet classification. After a packet has been classified, the packet needs to be marked by setting designated bits in the IP header. The following sections describe classification and marking:
•
•
•
•
•
Packet Classification Overview
To guarantee bandwidth for VoIP packets, a network device must be able to identify VoIP packets in all the IP traffic flowing through it. Network devices use the source and destination IP address in the IP header or the source and destination User Datagram Protocol (UDP) port numbers in the UDP header to identify VoIP packets. This identification and grouping process is called classification and it is the basis for providing any QoS.
Besides the static classification methods involving Layer 3 or Layer 4 header information matching, you can use a mechanism such as Resource Reservation Protocol (RSVP) for dynamic classification. RSVP uses H.245 signaling packets to determine which UDP port the voice conversation will use. It then sets up dynamic access lists to identify VoIP traffic and places the traffic into a reserved queue. RSVP is discussed later in this document.
Packet classification can be processor-intensive, so it should occur as far out toward the edge of the network as possible. Because every hop still needs to make a determination on the treatment a packet should receive, you need to have a simpler, more efficient classification method in the network core. This simpler classification is achieved through marking or setting the type of service (ToS) byte in the IP header.
The three most significant bits of the ToS byte are called the IP Precedence bits. Most applications and vendors currently support setting and recognizing these three bits. Marking is evolving so that the six most significant bits of the ToS byte, called the Differentiated Services Code Point (DSCP), can be used to define differentiated services (DS) classes. DSCP is discussed later is this document.
After every hop in the network is able to classify and identify the VoIP packets (either through port address information or through the ToS byte), those hops can then provide each VoIP packet with the required QoS. At that point, you can configure special techniques to provide priority queueing to make sure that large data packets do not interfere with voice data transmission, and to reduce bandwidth requirements by compressing the 40-byte IP plus UDP plus RTP header to 2 to 4 bytes.
Classification and Marking
Classification is the process of identifying the class or group to which a packet belongs. Network devices use various match criteria to place traffic into a certain number of classes. Matches are based on the following criteria:
•
•
•
•
•
•
As mentioned, it can be processor-intensive if nodes must repeat classification based on access list matches. Therefore, nodes should mark packets as soon as they have identified and classified the VoIP packets. If a node can set the IP Precedence or DSCP bits in the ToS byte of the IP header as soon as it identifies traffic as being VoIP traffic, then all of the other nodes in the network can classify based on these bits.
Marking is the process of the node setting one of the following:
•
•
•
•
•
In most IP networks, marking IP Precedence or DSCP should be sufficient to identify traffic as VoIP traffic.
Voice Dial Peers Classification and Marking Example
With Cisco VoIP gateways, you typically use voice dial peers to classify the VoIP packets and mark the IP Precedence bits. The following configuration example shows how to mark the IP Precedence bits:
Committed Access Rate Classification and Marking Example
Committed access rate (CAR) is an older technique that involves rate-limiting or policing traffic that matches certain criteria to an upper bound. CAR supports most of the matching mechanisms and allows IP Precedence or DSCP bits to be set differently depending on whether packets conform to or exceed a specified rate.
In general, CAR is more useful for data packets than for voice packets. For example, all data traffic coming in on an Ethernet interface at less than 1 Mbps can be placed into IP Precedence Class 3, and any traffic exceeding the 1 Mbps rate can go into Class 1 or be dropped. Other nodes in the network can then treat the exceeding or nonconforming traffic marked with lower IP Precedence differently. All voice traffic should conform to the specified rate if it has been provisioned correctly.
The following configuration example shows how to use CAR to classify and mark VoIP packets:
Policy-Based Routing Classification and Marking Example
Policy-based routing (PBR) is another older feature that allows traffic to be routed based on source port or access list. It also can be used to classify and mark packets. A simple configuration example follows:
Modular QoS Command-Line Interface Classification and Marking Example
The recommended classification and marking method is the Modular QoS Command-Line Interface (Mod QoS CLI, or MQC) feature, a template-based configuration method that separates the classification from the policy, allowing multiple QoS features to be configured together for multiple classes. You use a class map to classify traffic based on various match criteria and a policy map to determine what should happen to each class. Then you apply the policy to incoming or outgoing traffic on an interface using the service-policy interface configuration command. The following configuration example shows how to use Modular QoS to classify and mark packets:
QoS Queueing Mechanisms
After all traffic has been placed into QoS classes based on their QoS requirements, you need to provide bandwidth guarantees and priority servicing through an intelligent output queueing mechanism. This section describes queueing mechanisms and includes the following subsections:
•
Low Latency Queueing
A priority queue is required for VoIP. You can use any queueing mechanism that effectively gives VoIP high priority, but low latency queueing (LLQ) is recommended because it is flexible and easy to configure.
The most flexible queueing method that satisfies VoIP requirements is LLQ. LLQ uses the MQC configuration method to provide priority to certain classes and to provide guaranteed minimum bandwidth for other classes. During periods of congestion, the priority queue is policed at the configured rate so that the priority traffic does not monopolize all the available bandwidth. (If the priority traffic monopolizes the bandwidth, it prevents bandwidth guarantees for other classes from being met.) If you provision LLQ correctly, the traffic going into the priority queue should never exceed the configured rate.
LLQ also allows queue depths to be specified to determine when the router should drop packets if too many packets are waiting in any particular class queue. There is also a class default that is used to determine treatment of all traffic not classified by a configured class. The class default can be configured with the fair-queue interface configuration command, which means that each unclassified flow will be given an approximately equal share of the remaining bandwidth. Figure 1 shows how LLQ works.
Figure 1 LLQ Operation
In Figure 1, all traffic going out of an interface or subinterface (for Frame Relay and ATM) is first classified using MQC. There are four classes: one high priority class, two guaranteed bandwidth classes, and a default class. The priority class traffic is placed into a priority queue and the guaranteed bandwidth class traffic is placed into reserved queues. The default class traffic can be given a reserved queue or can be placed in an unreserved default queue where each flow will get an approximately equal share of the unreserved and available bandwidth. The scheduler services the queues so that the priority queue traffic is output first unless it exceeds a configured priority bandwidth and this bandwidth is needed by a reserved queue (that is, there is congestion). The reserved queues are serviced according to their reserved bandwidth, which the scheduler uses to calculate a weight. The weight is used to determine how often a reserved queue is serviced and how many bytes are serviced at a time. The scheduler services are based on the weighted fair queueing (WFQ) algorithm, a discussion of which is beyond the scope of this document.
If the priority queue fills up because the transmission rate of priority traffic is higher than the configured priority bandwidth, the packets at the end of the priority queue will be dropped only if no more unreserved bandwidth is available. None of the reserved queues are restricted to the configured bandwidth if bandwidth is available. Packets violating the guaranteed bandwidth and priority are dropped only during periods of congestion. You must therefore provision the priority queue with enough bandwidth to handle all the VoIP traffic requiring priority servicing.
LLQ Configuration Example
The following configuration example shows how to configure LLQ:
Other QoS Queueing Mechanisms
Several other queueing methods are available. For example, Modified Deficit Round Robin (MDRR) is a queueing mechanism available on the Cisco 12000 series Gigabit Switch Routers (GSRs) that allows bandwidth guarantees and priority servicing based on IP Precedence, DSCP, and MPLS EXP classes. MDRR supports one priority queue, seven reserved queues, and one multicast queue.
Once again, VoIP requires priority but there are data applications that cannot be starved and need bandwidth guarantees. You can use any queueing mechanism that effectively gives VoIP high priority, but we recommend LLQ.
Table 1 describes some of the available software queueing mechanisms.
Fragmentation and Interleaving
Because VoIP transmissions are extremely delay-sensitive, VoIP packets must be interleaved or inserted between data packet fragments. This section describes fragmentation and interleaving, and includes the following subsections:
•
•
•
Fragmentation and Interleaving Overview
Even if queueing is working at its best and prioritizing voice traffic, there are times when the priority queue is empty and a packet from another class is serviced. Packets from guaranteed bandwidth classes must be serviced according to their configured weight. If a priority voice packet arrives in the output queue while these packets are being serviced, the VoIP packet could wait a substantial amount of time before being sent. If we assume that a VoIP packet will need to wait behind one data packet, and that the data packet can be, at most, equal in size to the MTU (1500 bytes for serial and 4470 bytes for high-speed serial interfaces), we can calculate the wait time based on link speed.
For example, for a link speed of 64 kbps and MTU size of 1500 bytes, we have:
Serialization delay = (1500 bytes * 8 bits/byte) / (64,000 bits/sec) = 187.5 msTherefore, a VoIP packet may need to wait up to 187.5 ms before it can be sent if it gets delayed behind a single 1500-byte packet on a 64-kbps link. VoIP packets usually are sent every 20 ms. With an end-to-end delay budget of 150 ms and strict jitter requirements, a gap of more than 180 ms is unacceptable.
You need a mechanism that ensures that the size of one transmission unit is less than 10 ms. Any packets that have more than 10-ms serialization delay need to be fragmented into 10-ms chunks. A 10-ms chunk or fragment is the number of bytes that can be sent over the link in 10 ms. You can calculate the size by using the link speed:
Fragmentation size = (0.01 seconds * 64,000 bps) / (8 bits/byte) = 80 bytesIt takes 10 ms to send an 80-byte packet or fragment over a 64-kbps link.
On low speed links where a 10-ms-sized packet is smaller than the MTU, fragmentation is required. Simple fragmentation is insufficient, though, because if the VoIP packet must wait behind all the fragments of a single large data packet, the VoIP packet still will be delayed beyond the end-to-end delay limit. The VoIP packet must be interleaved or inserted in between the data packet fragments. Figure 2 illustrates fragmentation and interleaving.
Figure 2 VoIP Packet Fragmentation and Interleaving
Table 2 shows recommended fragment sizes for various link speeds based on the 10-ms rule.
Table 2 Link Speed and Fragmentation Size
Note
Three link fragmentation and interleaving (LFI) mechanisms are available. Table 3 lists their benefits and limitations.
MLP Link Fragmentation and Interleaving Example
The following configuration example shows how to configure fragmentation and interleaving using MLP LFI:
FRF.12 Fragmentation and Interleaving Example
The following configuration example shows how to configure fragmentation and interleaving using FRF.12:
Traffic Shaping
Traffic shaping is a QoS mechanism used to send traffic in short bursts at a configured transmission rate. It is most commonly used in Frame Relay environments where the interface clock rate is not the same as the guaranteed bandwidth or CIR. This section describes traffic shaping and includes the following subsections:
•
Traffic Shaping Overview
Frame Relay traffic shaping is the most common traffic shaping application in VoIP environments. Frame Relay scenarios usually have a hub and spoke network where the hub link speed is higher than any of the remote link speeds. In some cases, the sum of the remote link speeds is higher than the hub link speed, causing oversubscription. Without Frame Relay traffic shaping, the hub may try to send at higher rates than the remotes can receive traffic, causing the Frame Relay network to arbitrarily drop traffic. However, the remotes could all send at an aggregate rate that is higher than what the hub can receive, again causing the Frame Relay network to arbitrarily drop traffic. When we refer to the Frame Relay network, we mean the service provider (SP) network of WAN switches that provide the end-to-end PVC connectivity. Because the WAN SP cloud has no Layer 3 or above intelligence, it can drop VoIP traffic if contracts are violated. Therefore, you need to control transmission rates into a Frame Relay cloud so that you can control which packets get dropped and which packets receive priority servicing. Figure 3 shows an example of a typical Frame Relay network without traffic shaping.
Figure 3 Frame Relay Network
Frame Relay Traffic Shaping Example
The following configuration example shows how to configure Frame Relay traffic shaping:
IP RTP Header Compression
IP RTP header compression reduces the 40 byte IP+UDP+RTP header to 2 to 4 bytes, thereby reducing the bandwidth required per voice call on point-to-point links. The header is compressed at one end of the link and decompressed at the other end. Another standard name for this technique is cRTP, which stands for compressed RTP. Figure 4 shows the functionality of RTP header compression.
Figure 4 RTP Header Compression
To configure IP RTP header compression, you need to configure the ip rtp header-compression command under the serial interface, or the frame-relay ip rtp header-compression command under the Frame Relay subinterface. You can also configure the ip rtp compression-connections interface configuration command to set a maximum number of flows that will be compressed. Because cRTP can be processor intensive, you need to limit the number of compressed flows to prevent router performance degradation. Compressed RTP is recommended on low speed links where bandwidth is scarce and there are few VoIP calls.
Differentiated Services for VoIP
The Differentiated Services (DS) architecture QoS model provides a scalable mechanism to classify packets into groups or classes that have similar QoS requirements. This section describes DS and includes the following subsections:
•
•
•
DS and the DSCP (RFC 2474, RFC 2475) Overview
The first IP networks were based on the best-effort service model, which meant that delay, jitter, packet loss, and bandwidth allocation were unpredictable. Today a large number of networks still follow this best-effort model and do not support enhanced applications that require some sort of service guarantee.
Using the best-effort model, service providers have no means of offering service level agreements (SLAs) to their customers other than overprovisioning their network to deal with the busiest traffic hours. Enterprise customers and end users have no way of providing priority treatment or guaranteed bandwidth for VoIP. Traffic is treated on a simple, FIFO basis with no QoS enforcement.
The first architectural approach to providing end-to-end QoS required that the application signal its QoS resource requirements (such as bandwidth and guaranteed delay) to the network. In a VoIP scenario, this architectural approach meant that either the IP telephone or voice gateway needed to make QoS requests to every hop in the network so that end-to-end resources would be allocated. Every hop needed to maintain call state information to determine when to release the QoS resources for other calls and applications, and if enough resources were available, to accept calls with QoS guarantees. This method is called the Integrated Services QoS model. The most common implementation of Integrated Services uses Resource Reservation Protocol (RSVP). RSVP has some advantages, such as Call Admission Control (CAC), where a call can be rerouted by sending an appropriate signal to the originator if the network does not have the QoS resources available to support it. However, RSVP also suffers from some scalability issues; RSVP and those issues are discussed later in this document.
The DS architecture is the most widely deployed and supported QoS model today. It provides a scalable mechanism to classify packets into groups or classes that have similar QoS requirements and then gives these groups the required treatment at every hop in the network. The scalability comes from the fact that packets are classified at the edges of the DS "cloud" or region and marked appropriately so that the core routers in the cloud can provide QoS based simply on the DS class. The six most significant bits of the IP Type of Service (ToS) byte are used to specify the DS class; the Differentiated Services Code Point (DSCP) defines these six bits. The remaining two bits in the IP ToS byte are currently unused.
Figure 5 shows how the IP header defines the DS class.
Figure 5 Differentiated Services Field Definition
Differentiated Services is described and defined in the following RFCs:
•
•
•
•
RFC 2474 proposes a way of interpreting a field that has always been part of an IP packet. As mentioned previously in this document, the ToS field describes one entire byte (eight bits) of an IP packet. Precedence refers to the three most significant bits of the ToS byte; that is, [123]45678. (Occasionally the term ToS is used to refer to the next three bits: 123[456]78; however, in this document, to be consistent with the original RFC specification for the IP header (RFC 791), ToS refers to the entire set of eight bits.)
The first three bits of the DSCP are used as class selector bits; the class selector bits makes DSCP compatible with IP Precedence because IP Precedence uses the same three bits to determine class. Table 4 shows IP Precedence bit values mapped to DSCP.
Table 4 IP Precedence to DSCP Mapping
The next two bits are used to define drop preference. For example, if the traffic in Class 4 (the first three bits are 100) exceeds a certain contracted rate, the excess packets could be re-marked so that the drop preference is raised instead of being dropped. If congestion were to occur in the DS cloud, the first packets to be dropped would be the "high drop preference" packets. This is similar to DE-bit marking in Frame Relay and CLP-bit marking in ATM. These mechanisms allow the Layer 2 network to make intelligent drop decisions for nonconforming traffic during periods of congestion. DS allows similar operation over an IP network. The sixth bit must be set to 0 to indicate to the network devices that the classes have been set according to the DS standard.
The DS architecture defines a set of traffic conditioners that are used to limit traffic into a DS region and place it into appropriate DS classes. Meters, markers, shapers, and droppers are all traffic conditioners. Meters basically are policers, and class-based policing (which you configure using the police policy-map configuration command under a class in Modular QoS CLI) is a DS-compliant implementation of a meter. You can use class-based marking to set the DSCP and class-based shaping as the shaper. Weighted Random Early Detect (WRED) is a dropper mechanism that is supported, but you should not invoke WRED on the VoIP class. Per hop behavior (PHB) describes what a DS class should experience in terms of loss, delay, and jitter. A PHB determines how bandwidth is allocated, how traffic is restricted, and how packets are dropped during congestion.
Three PHBs are defined in DS based on the forwarding behavior required:
•
•
•
The Assured Forwarding (AF) standard specifies four guaranteed bandwidth classes and describes the treatment each should receive. It also specifies drop preference levels, resulting in a total of 12 possible AF classes, as shown in Table 5.
Table 5 Possible Assured Forwarding Classes
You would most likely use Assured Forwarding classes for data traffic that does not require priority treatment and is largely TCP-based. Expedited Forwarding more closely matches VoIP QoS requirements.
Implementing DS for VoIP: Expedited Forwarding PHB (RFC 2598)
Expedited Forwarding (EF) is intended for delay-sensitive applications that require guaranteed bandwidth. An EF marking guarantees priority service by reserving a certain minimum amount of bandwidth that can be used for high priority traffic. In EF, the egress rate (or configured priority bandwidth) must be greater than or equal to the sum of the ingress rates, so that there is no congestion for packets marked EF. You implement EF behavior by using the strict priority queue in low latency queueing (LLQ). Constant bandwidth is guaranteed for traffic belonging to the EF class, but at the same time if there is congestion, nonconforming packets exceeding the specified priority rate are dropped to assure that packets in other queues belonging to different classes are not starved of bandwidth. The recommended DSCP value for EF is 101110 (46). The first three bits of this EF value correspond to IP Precedence 5, which is the recommended ip precedence dial-peer configuration command setting for VoIP traffic. Therefore, if IP devices in the network can recognize IP Precedence or DSCP for classification and marking purposes, you can provision end-to-end QoS.
DSCP Class-Based Marking Configuration Example
The DS architecture specifies how to classify, mark, police, and shape traffic entering a DS region and how to treat different classes at every hop in the DS region. At the DS edge, all IP packets are marked with the appropriate DSCP so that QoS can be provided based on the DSCP inside the DS region. The following configuration example shows how to configure DSCP marking at the edge using class-based marking:
All queueing and other QoS parameters now can be set to match on DSCP in the rest of the DS region.
In the remaining sections of this document, we will match IP Precedence 5 traffic as VoIP and IP Precedence 3 traffic as HTTP (web traffic), with all other traffic going into the default class. Similarly, DSCP 46 could be used for VoIP and DSCP 26 for HTTP. We could use several other classification and marking mechanisms, but to maintain consistency and simplicity, we will use IP Precedence.
Resource Reservation Protocol
Resource Reservation Protocol (RSVP) is an implementation of the Integrated Services architecture for QoS (RFC 2205). When VoIP was introduced, RSVP was immediately seen as a key component that would provide admission control and QoS for VoIP flows. However, the way RSVP and H.323 were previously integrated provided neither admission control nor adequate QoS for voice flows. Several enhancements have been made to address these limitations, and RSVP can now be used to implement CAC and to signal a desired QoS that will provide good quality voice end-to-end, even in the presence of congestion.
In this section RSVP is described (in general), focusing on a particular subset of platforms, topologies, and protocols. We also assume that you are using H.323 as the session protocol for a VoIP gateway-based network. This section includes the following subsections:
•
•
RSVP Overview
The initial implementation of RSVP for VoIP had two limitations. The first was that CAC could not be implemented with RSVP because the reservation process was not synchronized with the voice call signaling. A call would proceed even if the RSVP reservation had failed or had not been completed. The second limitation was that a successful RSVP reservation might not provide good voice quality during periods of network congestion. RSVP created a reserved queue-per-traffic flow within the WFQ system and relied on that system to guarantee a bounded delay. However, WFQ was unable in some cases to provide an acceptable bounded delay for voice. RSVP needed to be able to use the priority queue in LLQ to guarantee a bounded delay that would not affect voice quality. In addition, RSVP was not supported on ATM or on shaped Frame Relay PVCs.
You should deploy RSVP to improve VoIP QoS only where it can have a positive impact on quality and functionality. The benefits of using RSVP outweigh the costs (management, overhead, and performance impact) only where there is limited bandwidth and frequent network congestion. Some IP environments have enough bandwidth to guarantee the appropriate QoS without needing to implement CAC for every call.
The following four mechanisms were introduced in Cisco IOS software to handle resource-based CAC:
•
•
•
•
In this document we discuss only the use of RSVP for CAC.
RSVP for CAC Overview
Using RSVP for VoIP CAC requires the synchronization of the call setup signaling and the RSVP signaling. This synchronization guarantees that the called-party phone rings only after the resources for the call have been reserved. This synchronization also gives voice gateways the control of what action to take before the call setup moves to the alerting stage if the reservation fails or cannot be completed within a predefined period of time. A voice call will trigger two RSVP reservations because the reservation and admission control mechanisms provided by RSVP are unidirectional. Each voice gateway is responsible for initiating and maintaining one reservation toward the other voice gateway. CAC for a VoIP call fails if at least one of the reservations fails. Figure 6 shows the sequence of packets exchanged between the gateways during a call setup if RSVP is used for resource reservation.
Figure 6 Call Setup with RSVP Enabled
In Figure 6, an originating gateway (OGW) initiates a call toward a terminating gateway (TGW). The originating gateway sends an H.323 SETUP message to the terminating gateway to initiate the call. That SETUP message carries the QoS that the originating gateway considers acceptable for the call. The terminating gateway responds with an H.323 CALL PROCEEDING message. Both the originating gateway and the terminating gateway initiate a reservation request by sending an RSVP PATH message. The packet flows of both reservations are independent of each other unless one of them fails. The terminating gateway blocks the call setup process waiting for the reservation results. The terminating gateway controls the admission decision for the call and needs to be notified that the reservations in both directions were successful. The terminating gateway discovers that its reservation was successful when it receives the RSVP RESV message. The terminating gateway detects that the originating gateway reservation was successful when it receives an RSVP RESV CONFIRMATION message from the originating gateway. At this point, the terminating gateway lets the call setup continue and sends an H.323 ALERTING message to the originating gateway once it is notified that the called side is in alerting state. A normal disconnect is initiated when an H.323 RELEASE COMPLETE message is sent after the call is connected. At that point, the gateways tear down their reservations by sending RSVP PATH TEAR and RESV TEAR messages.
If at least one RSVP reservation fails, you can configure a voice gateway to take the following actions:
•
•
•
This last behavior is possible because the terminating gateway knows which QoS is acceptable for the call from its own configuration and the value included by the originating gateway in the H.323 SETUP message. If the terminating gateway and the originating gateway request a nonbest-effort QoS and at least one reservation fails, the call will proceed as best-effort only if the originating gateway and the terminating gateway are willing to accept best-effort service. Call release and call rerouting are possible if one of the two voice gateways will not accept best-effort service. If you configure the gateway to reject the call and report the failure, CAS trunks and analog lines generate a fast busy signal. On CCS PRI trunks, a Q.931 DISCONNECT message with a cause "QoS unavailable" (49) will be generated.
Figure 7 shows the details of a call that is rejected because the reservation initiated from the terminating gateway failed.
Figure 7 Call Failing RSVP CAC Because of Terminating Gateway Reservation Failure
Deploying CAC Based on RSVP
As already mentioned, you should deploy RSVP to improve VoIP QoS only where it can have a positive impact on quality and functionality. The benefits of using RSVP outweigh the costs only where bandwidth is limited. We recommend using Cisco IOS Release 12.1(5)T or a later release if you want to implement CAC for VoIP using RSVP.
You need to complete the following three basic steps to configure CAC for VoIP calls using RSVP:
•
•
•
The following configuration example shows how to configure CAC for VoIP calls using RSVP:
The default dial-peer configuration requests and accepts best-effort QoS for VoIP calls. This translates to the gateway not initiating an RSVP reservation for the call because IP provides best-effort service by default. The other two service alternatives are controlled-load or guaranteed-delay QoS. These two services require RSVP signaling; they are requested using the req-qos dial-peer configuration command. The acceptable QoS controls how strict or loose the CAC criteria should be; you configure the acceptable QoS controls by using the acc-qos dial-peer configuration command. We recommend that you configure the originating gateway and the terminating gateway to request and accept guaranteed delay.
Sometimes you can configure the implicit dial peer matched on a terminating gateway to request and accept best-effort QoS. This dial peer takes effect when there is not an explicit dial peer match.
Configuring Local Gateway Resources if CAC Fails
As mentioned earlier, you can configure a voice gateway to take different actions if admission control fails. The first alternative is to have the gateways signal the user or the switch that delivered the call with a fast busy signal or a disconnect cause. If the call was delivered to the gateway by an ISDN switch, you can tune the Q.931 disconnect cause to guarantee that the switch handles calls correctly. A "QoS unavailable" (49) cause is returned by default when an ISDN call fails CAC because of the requested and acceptable QoS configured. You can modify this cause with the isdn network-failure-cause interface configuration command or isdn disconnect-cause interface configuration command. The current implementation of the isdn network-failure-cause command overrides the value configured using the isdn disconnect-cause command.
The following configuration example shows how to configure local gateway resources if CAC fails by tuning the Q.931 Disconnect Cause:
A second option is to allow the gateway to reroute the call through another path. If the dial peer matched by the call is part of a hunt group, other dial peers in that group are tried according to the preference dial-peer configuration command. This allows you to implement different types of call routing on the gateway that consider QoS across IP networks.
The following configuration example shows how to configure local gateway resources by rerouting calls on the gateway if CAC fails:
The third possibility, available in Cisco IOS releases later than 12.1(5)T, is to configure the gateways to proceed with the call even if RSVP reservations fail. This option, however, does not provide a major improvement over earlier Cisco IOS release functionality. The only benefit it provides is that, in case of a successful RSVP reservation, the call does not proceed until the reservation is established.
As mentioned earlier, a call can fail admission control if at least one of the two RSVP reservations needed for the call fails. For each RSVP reservation, admission control is performed on all interfaces where you have enabled RSVP by using the ip rsvp bandwidth interface configuration command. You can configure two values with the ip rsvp bandwidth command: the maximum total reserved bandwidth and the maximum bandwidth per reservation. The maximum total bandwidth is limited by default to no more than 75 percent of the total bandwidth of the interface. You can modify that limit with the max-reserved-bandwidth interface configuration command. Exceptions to the maximum total bandwidth limitation are Frame Relay and ATM PVCs. For Frame Relay PVCs, the maximum reservable bandwidth is the minimum CIR, or, if not configured, half of the CIR. For ATM PVCs, the maximum reservable bandwidth is 75 percent of the configured available bit rate output minimum cell rate, near real-time VBR output-SCR, or real-time VBR average rate, whichever is configured. The total bandwidth available for RSVP reservations may be lower if you have reserved bandwidth using CBWFQ or LLQ through MQC. A bandwidth manager makes sure that the interface or the PVC bandwidth is not oversubscribed during the router operation.
Note
You should configure the maximum bandwidth per reservation to be no lower than what the codec requires plus all other protocol overhead except the Layer 2 protocol overhead. Table 6 shows the lowest values that you can use for different codecs. Remember that these values do not account for the bandwidth savings introduced by cRTP or voice activity detection (VAD). The actual voice stream may use less bandwidth, but the system will use the worst-case bandwidth.
One consideration when deploying RSVP for VoIP is the impact of resource reservation on the postdial delay. Implementing VoIP CAC based on RSVP relies on a prompt confirmation or rejection of the requested reservation. The time taken to reserve resources adds to the postdial delay, which should be kept as low as possible in most cases. RSVP packets are carried inside IP datagrams and are unreliable by nature. If an RSVP packet is lost during the initial reservation setup, an RSVP refresh timer must expire before the lost packet is resent. Because this refresh timer typically is defined in tens of seconds, a scenario that may add a postdial delay is unacceptable for the user. The call rsvp-sync resv-timer global configuration command lets you control the maximum amount of time that the terminating gateway waits for the result of RSVP reservation requests. The default value of this timer is 10 seconds; you can set it to a value from 1 to 60 seconds according to your expectation of postdial delay.
Using RSVP with LLQ
Flows requesting a particular QoS via RSVP can take advantage of the queueing alternatives available in LLQ, which has two major components: a strict priority queue (PQ) and a CBWFQ system. Earlier implementations of RSVP relied on WFQ to meet the QoS requirements for delay-sensitive traffic. A reserved queue with a low weight was created when the RSVP reservation was installed. However, WFQ could not meet the delay requirements of voice traffic, and voice calls using RSVP were not able to take advantage of the PQ available throughout LLQ.
In Cisco IOS Release 12.1(3)T and later releases, a priority profile based on traffic characteristics exists so that certain flows can take advantage of the strict PQ in LLQ. When a RSVP reservation request is received on an interface where you have enabled WFQ, the flow traffic specification (TSpec) is compared against the profile to decide if that flow should take advantage of the PQ or if a queue should be reserved on the WFQ system. The TSpec is the traffic description carried in RSVP messages. This traffic description is made in terms of a token bucket (token rate r, plus a bucket size b) and some additional parameters (peak rate p, minimum policed unit m, and maximum packet size M). The PQ profile is defined in terms of token rate, bucket size, and an optional peak rate to token rate ratio. Flow reservations with a TSpec that do not exceed those defined in the PQ profile will use the PQ. Those flows with a TSpec that exceeds at least one parameter defined in the profile will get a reserved queue in the WFQ system. The priority profile allows you to classify priority flows based on their traffic characteristics—not just on the transport protocol and port. Figure 8 shows the LLQ structure for an interface where traffic is classified into different queues using several methods, including RSVP.
Figure 8 RSVP Support for LLQ on Point-to-Point Interfaces
Cisco IOS Release 12.1(5)T introduced RSVP support for LLQ on Frame Relay PVCs. In this case, each PVC has its own queueing structure with a PQ and a CBWFQ system. At the interface level, a FIFO queue is set up unless you have enabled FRF.12 fragmentation. In that case, a dual FIFO system is set up with a high priority queue and a low priority queue. The high priority queue receives the PQ traffic from all PVCs plus Layer 2 control traffic. The low priority queue receives all other traffic from all PVCs. Remember that Frame Relay traffic shaping (FRTS) is required for Frame Relay circuits whether FRF.12 fragmentation is enabled or not. FRTS provides the back-pressure mechanism to detect congestion per PVC. Support for ATM PVCs is available in Cisco IOS Release 12.2(1)T.
Deploying RSVP Support for LLQ
You enable RSVP support for LLQ by default for voice flows on interfaces where RSVP and WFQ are enabled. You need not explicitly configure priority queues for voice packets. You can configure a custom priority queue profile using the ip rsvp pq-profile global configuration command. Configuring the profile as ip rsvp pq-profile voice-like restores the default behavior. The default priority queue profile uses a token rate of 12,288 bytes per second (approximately 98 kbps), a bucket size of 592 bytes, and a peak rate equal to 110 percent of the token rate (13,516 bytes per second or approximately 108 kbps). These parameter values support all possible codec configurations on voice gateways running Cisco IOS software. A Cisco voice gateway configured to reserve resources via RSVP will infer the correct TSpec exclusively from the codec used on the dial peer. You cannot control TSpec values using the CLI and no other bandwidth-saving features (such as VAD) are taken into consideration. Some revisions of Microsoft NetMeeting for Windows 98 and Windows 2000 (which use RSVP) signal a bucket size in the TSpec that is not compatible with these defaults. This problem affects Microsoft NetMeeting for calls using codecs that require 32 kbps or more. In those cases, you need to create a custom profile to match the parameters signaled by Microsoft Windows.
The following configuration example shows how to configure RSVP support for LLQ on a Frame Relay circuit that has two PVCs:
One of the important implications of RSVP support for LLQ is that it lets you classify voice traffic based on its traffic characteristics rather than on the transport protocol (UDP) and port number (16384 through 32767). The proper operation of LLQ relies on the assumption that the priority queue is used only by well-behaved traffic (such as voice) that has a predictable rate and a very low burst size. Classification based on transport protocol and ports could allow bursty or noncritical traffic into the priority queue, which might affect the quality of existing voice calls and the performance of the traffic using the WFQ system. You need to take the effects of bursty or noncritical traffic into account when you are defining a custom priority queue profile. You should understand all the implications on other types of traffic—in particular, when the PQ profile could let flows with some degree of burstiness into the priority queue. RSVP support for LLQ prioritizes voice packets but it does not take care of the voice signaling. It may not be possible to initiate new calls during periods of heavy congestion due to loss of signaling packets. To address this situation, you can reserve some amount of bandwidth explicitly for signaling packets using the MQC. You can also mark RSVP messages for special treatment using the ip rsvp signaling dscp interface configuration command.
In the following configuration example, voice packets are prioritized using RSVP; and the signaling is guaranteed a minimum bandwidth during periods of congestion through the MQC:
VoIP QoS over Leased Lines (Using PPP)
This section describes how to configure VoIP in a typical network where low speed WAN links are used to carry both data and voice traffic. It includes the following subsections:
•
•
VoIP over Leased Line (Using PPP) Scenario
A typical application of VoIP is for a large corporation to use its existing WAN infrastructure for data traffic to carry voice calls between its headquarters and its branch offices. Figure 9 shows a typical VoIP network environment where low speed WAN links are being used to carry both data and voice traffic.
Figure 9 Typical VoIP Network Environment
For low speed WAN links that are not well-provisioned to serve voice traffic, problems such as delay, jitter, and loss become even more pronounced. In this particular network environment, the following factors can contribute to poor voice quality:
•
•
•
•
•
–
–
In addition, each link and piece of equipment in the VoIP path adds delay to voice packet transmission. The possibility of voice packet loss also increases as voice traffic travels a longer distance and over more hops in the network. Low speed WAN connections usually are the weakest links.
VoIP over Leased Line (Using PPP) Recommended Solution
Under normal conditions, network equipment and end stations cannot differentiate between the requirements of real-time voice packets and standard data traffic. This could result in serious speech degradation. To ensure voice quality, you must classify data and voice traffic into different categories and give voice traffic priority handling across a shared data network backbone. Giving voice traffic priority handling minimizes delays and drops, and whenever possible, gives voice traffic predictable transmission performance. For PPP links, we recommend the following QoS features:
•
•
•
•
•
The following configuration example shows a complete configuration with all of the listed QoS features enabled:
VoIP QoS over Frame Relay Networks
This section describes how to configure VoIP in a typical network where Frame Relay WAN links are used to carry voice traffic. It includes the following subsections:
•
•
VoIP QoS over Frame Relay Scenario
Another typical VoIP application is for a large corporation to use its existing Frame Relay WAN data traffic infrastructure to carry voice calls between its headquarters and its branch offices. There are two options here: either carry the voice and data on separate PVCs or use the same PVC for voice and data traffic. In the first scenario, you must still give the voice traffic priority by using a technique such as PVC Interface Priority Queue (PIPQ). PIPQ lets you assign different priorities for PVCs—high, medium, normal, or low. PIPQ also allows PVCs to be queued at the main physical interface so that high priority traffic goes before medium, normal, and low priority traffic. PIPQ, however, has the same problem as priority queueing—the high priority traffic can starve the other traffic of bandwidth. However, if you use Frame Relay traffic shaping correctly, you can minimize this problem because each PVC will have a defined maximum transmission rate.
In the most common scenario, you use a single PVC to carry all the traffic between sites, as shown in Figure 10.
Figure 10 VoIP QoS over Low Speed Frame Relay Links
VoIP QoS over Frame Relay Recommended Solution
You need to configure Frame Relay traffic shaping to ensure that speed mismatches at the remote and hub sites are handled correctly. For example, if the hub site has a T1 connection into the Frame Relay network and the remote site has a 128-kbps access speed, the hub site has the capability to send at T1 speeds toward this single remote. The Frame Relay switches will buffer this traffic to a small extent, but then arbitrarily drop anything over 128 kbps. You need to decide what should be dropped and what should be prioritized at the endpoints of the PVC.
Frame Relay traffic shaping allows the routers to send traffic into the Frame Relay cloud below a preconfigured rate. Any traffic over this rate is queued, and a queueing algorithm such as LLQ can be used to make intelligent decisions on which packets should be sent. If the queues fill up, the packets simply will be dropped. However, if VoIP is given priority, and the total VoIP traffic is below the traffic-shaping rate, VoIP packets will be serviced with low latency and will not be dropped.
For lower speed links less than 1.2 Mbps, you need to configure packet fragmentation to ensure that a VoIP packet need not wait behind a large packet. Fragmenting larger data packets to 10 ms of the link speed can bind the maximum waiting period. You can use cRTP to efficiently use bandwidth if the number of calls is not too large.
To provide high quality to VoIP over Frame Relay, you need to configure the following features:
•
–
–
–
•
•
•
The following configuration example shows how to enable the QoS features described in the previous section:
VoIP QoS over ATM
This section describes how to configure VoIP QoS over ATM and includes the following subsections:
•
•
VoIP QoS over ATM Scenario
ATM technology has inherent advantages in handling VoIP traffic because of its small, fixed-size cells and class of service (CoS) mechanisms. These advantages do not ensure, however, that VoIP traffic will automatically obtain the QoS it needs from the ATM network carrying it. VoIP traffic will not automatically obtain the QoS it needs because QoS definitions at the IP layer, such as the IP Precedence settings in the packet header, do not automatically match ATM CoS settings, namely traffic class (Constant Bit Rate (CBR), Variable Bit Rate (VBR), Available Bit Rate (ABR), Undefined Bit Rate (UBR)) and traffic parameters such as Sustainable Cell Rate (SCR), Peak Cell Rate (PCR), and burst size. Consequently, after data and voice packets are identified and sorted at the IP layer, the network operator must manually configure the ATM virtual circuits (VCs) to ensure QoS for voice packets across an ATM network. This manual provisioning is time-consuming, labor-intensive, error-prone, and, above all, does not scale as more and more voice traffic is introduced into the network.
Figure 11 shows an example of VoIP QoS configured to support ATM.
Figure 11 VoIP QoS over ATM Links
Two solutions are available for providing QoS to VoIP over an ATM network: one using separate data and voice VCs and one using shared data and voice VCs.
VoIP QoS over ATM Solution Using Separate Data and Voice ATM PVCs
For data and voice traffic sharing the same destination but requiring different QoS, you need to define groups of ATM VCs to form PVC bundles. In a PVC bundle, all the PVCs share the same source and destination, and each bundle is assigned to carry IP traffic with a specific IP Precedence level or range of levels. After you configure PVC bundles, you then configure each PVC with its specific ATM QoS parameters. As voice and data traffic with different IP Precedence levels arrives at the ATM interface of the router, Cisco IOS software dynamically sends it out on the appropriate PVC, effectively mapping IP QoS classes to ATM CoSs.
The key benefits of implementing VoIP QoS using this method are as follows:
•
•
The following configuration example shows how to configure VoIP over ATM using PVC bundles to separate voice and data PVCs:
VoIP QoS over ATM Solution Using Shared Data and Voice ATM PVCs
If you decide to use separate PVCs for voice and data, you must adjust the bandwidth allocation accordingly as voice traffic grows beyond the bandwidth configured on the voice PVC. This manual reprovisioning is not necessary when voice and data share the same PVC, provided that voice always gets the priority it needs. You can configure VoIP traffic to have absolute priority over data traffic by configuring LLQ on the ATM PVC.
The following configuration example shows how to configure VoIP over ATM using the same PVC for data and voice traffic:
Related Documents
•
http://www.cisco.com/univercd/cc/td/doc/product/software/ios121/121cgcr/qos_c/index.htm•
http://www.cisco.com/univercd/cc/td/doc/product/software/ios121/121cgcr/qos_r/index.htm•
http://www.cisco.com/univercd/cc/td/doc/product/software/ios121/121cgcr/multi_c/index.htm•
http://www.cisco.com/univercd/cc/td/doc/product/software/ios121/121cgcr/multi_r/index.htm•
http://www.cisco.com/univercd/cc/td/doc/product/software/ios121/121newft/121t/121t1/
copsrsvp.htm•
http://www.cisco.com/univercd/cc/td/doc/product/software/ios121/121newft/121t/121t2/cbpmark.htm•
http://www.cisco.com/univercd/cc/td/doc/product/software/ios121/121newft/121t/121t2/
clsbsshp.htm•
http://www.cisco.com/univercd/cc/td/doc/product/software/ios121/121newft/121t/121t2/
dtfrhccc.htm•
http://www.cisco.com/univercd/cc/td/doc/product/software/ios121/121newft/121t/121t2/
dtfrpqfq.htm•
http://www.cisco.com/univercd/cc/td/doc/product/software/ios121/121newft/121t/121t3/
dtcfgbst.htm•
http://www.cisco.com/univercd/cc/td/doc/product/software/ios121/121newft/121t/121t3/
rsvp_llq.htm•
http://www.cisco.com/univercd/cc/td/doc/product/software/ios121/121newft/121t/121t5/
cbpmark2.htm•
http://www.cisco.com/univercd/cc/td/doc/product/software/ios121/121newft/121t/121t5/
dtcbwred.htm•
http://www.cisco.com/univercd/cc/td/doc/product/software/ios121/121newft/121t/121t5/
dtdcrtp.htm•
http://www.cisco.com/univercd/cc/td/doc/product/software/ios121/121newft/121t/121t5/
dtllqvip.htm•
http://www.cisco.com/univercd/cc/td/doc/product/software/ios121/121newft/121t/121t5/
dtdts.htm•
http://www.cisco.com/univercd/cc/td/doc/product/software/ios121/121newft/121t/121t5/
dtdfsv.htm•
http://www.cisco.com/univercd/cc/td/doc/product/software/ios121/121newft/121t/121t5/
dtlfifra.htm•
http://www.cisco.com/univercd/cc/td/doc/product/software/ios121/121newft/121t/121t5/
dtpoli.htm•
http://www.cisco.com/univercd/cc/td/doc/product/software/ios121/121newft/121t/121t5/
dt4trsvp.htm
Feedback