|
|
Table Of Contents
Solution Transport Architecture
Aggregation and Distribution Transport Architecture
Configuration 1: 10-GE Layer 3 Ring
Configuration 2: 1-GE plus 10-GE Hub and Spoke
Overview of DiffServ Architecture
DiffServ Architecture in the Solution
QoS in the Aggregation/Distribution Network
Solution Transport Architecture
The Cisco Wireline Video/IPTV Solution transport architecture is subdivided into recommendations for the access, aggregation, and distribution networks. While the service mapping architecture used in Release 1.1 includes requirements regarding how a residential gateway (RG) interfaces to the home network, it does not include any recommendations for the technologies or configurations used in that network. The Release 1.1 transport architecture also does not include recommendations for the core network. Because of this, the solution test environment combines the video application components of both the super headend (SHE) and video headend office (VHO) sites into a single combined topology that connects aggregation routers (ARs) to a distribution edge router (DER).
This chapter presents the following major topics:
•
Overview
•
Overview
Figure 3-1 illustrates the transport layers of the general VoD transport architecture described in Common Broadcast Video and VoD Components. These layers are the subject of the recommendations in this document.
While the solution transport architecture focuses on video, there is an implicit assumption that the network be able to support a full triple-play environment. Consequently, the transport architecture includes a common quality of service (QoS) architecture for video, voice, and Internet access services. Because the transport architecture is based on the service mapping model described in Video Transport Architecture and Issues, the actual transport architecture used for Internet access and voice services may differ from the video transport architecture described in this document.
To ensure that the video transport architecture works in a triple-play environment, solution testing included a test bed environment in which the transport network was configured to support all three services. Because solution testing was focused on video, it included the application, control, and transport environment for video services. Testing only included enough testing of the Internet access and voice services to ensure that an example forwarding architecture for these services can coexist with video, and that the common QoS architecture specified in this document meets the jitter and packet-loss requirements for each service.
Figure 3-1 IPTV/Video over Broadband Transport Architecture: Solution Core
Note
While the transport architecture includes configuration recommendations for all of the transport layers shown in Figure 3-1, this document includes example configurations only for the transport components that are implemented by means of Cisco products. These components are the DER and AR. Because the DER and ARs are the switching components that implement the distribution and aggregation networks, more detailed configuration information is provided for this portion of the network.
Note
Solution Components
Table 3-1 lists the network architecture components used in Release 1.1 of the solution, with additional information. For detail regarding interfaces, see Table 3-5.
Aggregation and Distribution Transport Architecture
As described in Video Transport Architecture and Issues, the transport architecture uses service mapping to support the capability of having separate routing and forwarding planes for different services. This functionality is used in the aggregation and distribution networks to enable a separate logical and physical transport architecture that is optimized for the delivery of video.
This section describes how the transport architecture is optimized for video. Because an important requirement of the transport architecture is that it also must support a triple-play environment, this section also describes an example distribution and aggregation network configuration for voice and Internet access.
This section presents the following topics:
Video Forwarding
This section presents the following topics related to the delivery of video services:
•
•
Layer 3 Edge for Video Services
One of the primary architectural decisions that must be made in specifying a transport architecture for video services is where the Layer 3 edge of the transport network should be for video services. Figure 3-2 illustrates the points in the network where the Layer 3 edge may reside, as well as the issues and benefits associated with each location. There are three points: the DSLAM, the AR, or the DER. This section describes the issues and benefits associated with each of these options, as well as the design choice that was made for Release 1.1.
Figure 3-2 Potential Layer 3 Edge Points for Video
Table 3-2 summarizes issues and benefits for edge points 1, 2, and 3 in the above figure. The paragraphs that follow address issues to related to DSLAM-based, DER-based, and AR-based Layer 3 edge points.
DSLAM-Based Layer 3 Edge
Because of their location at the edge of the network, DSLAMs have traditionally performed Layer 2 switching functions. This has kept the function of the DSLAM fairly simple and has also made DSLAMs simple to manage. However, a Layer 3-capable DSLAM is more complex to build, and therefore more complex to manage.
Issue: DSLAM Complexity
A DSLAM that supports Layer 3 functionality must be capable of a number of functions besides Layer 3 forwarding. For example, a Layer 3-capable DSLAM must be able to support a DHCP relay function. This function requires that the IP address of a DHCP server as well as the IP subnet that the DSLAM is associated with must be configured on the DSLAM. The DSLAM must also support and be configured for IP routing protocols to enable dynamic routing from the AR.
Issue: Complex Subscriber-Address Management
An IP-capable DSLAM must have an IP subnet allocated to it to allow IP packets to be routed to it. This complicates IP address management, because a separate IP subnet must be allocated for each DSLAM. This also makes IP address management for the residence more complex, as separate IP address pools must be allocated for each DSLAM.
DER-Based Layer 3 Edge
The DER may also be at the Layer 3 edge for video. With this type of design, forwarding in both the aggregation and distribution networks is performed at Layer 2. While such a design is consistent with common designs for PPPoE-based Internet access services, it creates a number of scaling issues for both the ARs and the DER. This design can also create issues for video services, because of the flooding associated with common learning-bridge architectures.
Issue: Scaling for the Layer 2 MAC Table and Layer 3 Forwarding Table
To understand the scaling issues associated with this design, it is useful to look at the number of STBs that may be aggregate by a single DER. To provide worst-case scaling numbers, we use the following numbers for a hypothetical VoBB deployment:
•
•
•
Therefore, in this example, the DER is aggregating 160,000 subscribers.
When the DER is configured as the Layer 3 edge for video services, all STBs that are connected through that router are in the same IP subnet. If the subnet is aggregated as a single Layer 2 topology, each of the ARs aggregated by the DER need to support MAC table forwarding entries for all of the STBs on that subnet. This amounts to 160,000 MAC table entries for each AR. This requirement drives up the cost of ARs, because each MAC table entry requires a hardware content-addressable memory (CAM) table entry. There are methods that use separate VLANs to divide the distribution layer topology into simpler Layer 2 topologies that are aggregated at the DER. These methods reduce the MAC table scaling requirements for the ARs, but result in a more complex Layer 2 topology to administer in the distribution network.
Another issue with configuring the DER as the Layer 3 edge for video services is that this router must maintain a separate ARP table entry and forwarding table adjacency for each STB aggregated through it. In our previous example, this amounts to 160,000 such adjacencies. Such a large number again results in higher cost for this router, because each forwarding adjacency uses a separate hardware ternary CAM (TCAM) entry. By comparison, if the Layer 3 edge in the example above were moved to the AR, this device would need to support only 16,000 ARP table entries and forwarding adjacencies.
Issue: Multicast Configuration Complexity and Transport Inefficiencies
A network design that aggregates multicast video traffic at Layer 2 results in a complex multicast configuration, as well as in significant inefficiencies in multicast traffic behavior. Figure 3-3 illustrates the configuration complexity and transport inefficiencies when a Layer 2 distribution network is used for multicast video.
Figure 3-3 Multicast Convergence with Layer 2 Distribution
When multicast video is aggregated at Layer 2, the resulting design typically uses more than one DER for redundancy. As a result, the PIM protocol state machine elects a designated router (DR). The DR is responsible for registering sources and sending upstream join and prunes on behalf of the members of the subnet (VLAN). In addition, the network selects an IGMP querier for the served subnet. The IGMP querier is responsible for sending IGMP queries on the subnet served by the redundant IP edge routers.
Each aggregation switch (AS) is responsible for replicating multicast streams from the distribution network to aggregation ports that have subscribers joined to them. As shown in Figure 3-3, there are two potential sources inserting video into the distribution network. These are DER 1 and DER 2. Because either of these two sources may be used to send multicast traffic onto the ring, each aggregation switch must send IGMP joins up both of the uplinks. This IGMP behavior makes it very difficult for the Layer 2 switches to determine when and when not to replicate multicast traffic on the distribution ring. To make multicast work properly in this type of environment, each port on each switch must be configured to replicate packets dynamically by using IGMP or statically. Ports that are configured to replicate dynamically send the traffic associated with a multicast group only if there has been an IGMP join issued for that multicast group. Ports that are configured to replicate statically send all multicast traffic all the time, independently of whether an IGMP join has been issued. In the case of Figure 3-3, each upstream port on each switch must be configured for static replication, because the downstream multicast traffic could potentially flow from either direction on the ring. This configuration results in additional complexity when multicast is configured on redundant topologies.
In addition to being more complex to configure in a redundant topology, multicast is less efficient. This is because multicast streams must be sent everywhere in the Layer 2 ring, independently of where an IGMP join was issued. Figure 3-3 illustrates an example multicast replication in a Layer 2 environment. Here the subscriber has issued a channel-change request from the STB attached to AR 2. The channel-change request results in an IGMP join message being propagated in both directions of the distribution network to both DER 1 and DER 2. DER 2 has been elected as the designated router, so it translates the IGMP join into a PIM join, while DER 1 ignores the IGMP join request. As a result of the IGMP join, DER 2 sends the multicast stream to the ring. Because AS 2 is using IGMP snooping on the downstream link, it is the only switch that replicates the stream to the DSLAM. Note, however, that the multicast traffic gets propagated all the way through the Layer 2 ring to DER 1. Each switch must replicate the traffic to other switches on the ring, because it is very difficult to determine where to send the multicast traffic on the ring based on IGMP snooping alone. DER 1 drops the multicast traffic when it receives it, because it does not have any "downstream" requestors for the stream. The result of using Layer 2 in the distribution network is that bandwidth is wasted on the distribution ring, because the multicast stream must be sent everywhere—independently of which nodes on the ring have asked for the traffic.
Figure 3-4 illustrates multicast operation and traffic flow when a Layer 3 distribution network is used for video. Here all nodes in the redundant topology are in the same Layer 3 topology. This results in simpler configuration as well as a more efficient traffic flow pattern. IP multicast is inherently different from Layer 2 forms of replication, because the multicast tree is built from PIM messages that are routed from the edge of the IP network to the source by means of reverse-path routing. Reverse-path routing is essentially the same as destination-based routing, except that the path to the source is looked up on the basis of the IP source address. This figure illustrates how the PIM messages are routed to the source and how the multicast distribution tree is built more efficiently as a result.
Figure 3-4 Multicast Traffic Flow with Layer 3 Distribution
In this figure, the subscriber has again issued a channel-change request from the STB attached to AR 2. The request results in an IGMP join message being sent to DER 2. Release 1.1 of the solution uses Source Specific Multicast (SSM), along with SSM mapping, as the IP multicast technology for the broadcast video service. As a result, AR 2 can translate the IGMP join request into the IP address of the encoder that is being used to generate that stream. With the IP source address, AR 2 uses reverse-path routing to decide where to send an PIM message. In this case, the shortest path to the primary source is through DER 2.
Once PIM state is established, DER 2 replicates the multicast stream to AR 2, which in turn sends the multicast stream to the DSLAM and the STB. Note that the multicast stream is not replicated throughout the distribution ring as it was in the Layer 2 scenario. This is because reverse-path route lookup results in a multicast tree that is built from the source directly to the nodes that requested the traffic. The result of using Layer 3 in the distribution network is an IP multicast environment that is simpler to configuration and more efficient in bandwidth use than are Layer 2 environments.
AR-Based Layer 3 Edge
When the AR is configured as the Layer 3 edge for video, the network is typically configured so that the AR is located at a different point in the network than the Layer 3 edge for Internet access services. This type of configuration may be considered not as architecturally "clean" as having the Layer 3 edge for all services located at the same point in the network. However, as described below, these issues are far outweighed by the benefits of using a Layer 3 distribution network for video services. Release 1.1 of the solution uses an AR-based Layer 3 edge to take advantage of these benefits.
Note that the AR may not be the node that directly aggregates the GE uplinks from DSLAMs. The AR is defined as the first node in the physical topology that aggregates enough subscribers that either path or node redundancy is required for video services. In a ring topology, the AR is defined as the node that connects the ring to a nonredundant hub-and-spoke aggregation architecture. In a hub-and-spoke topology, the AR is defined as the first node that includes redundant uplinks to the distribution network. In topologies where the AR does not terminate the GE uplinks from DSLAMs, there may be a Layer 2 aggregation network between the DSLAMs and the AR that does not include either path or node redundancy. Layer 2 Aggregation Alternatives, provides details on Layer 2 aggregation schemes that may be used between DSLAMs and the AR.
The sections below provide details on some of the benefits that make the AR the best choice as the Layer 3 edge in a video topology.
Benefit: Source-Specific Multicast
When the AR is configured as the Layer 3 edge for video services, the distribution network can take advantage of IP multicast features such as Source Specific Multicast (SSM). SSM is a technology that enables the network to build a separate distribution tree for each multicast source. SSM simplifies the operational complexity of configuring a multicast network, because it does not require the configuration of a rendezvous point (RP) to allow multicast forwarding as non-source-specific multicast technologies do. In addition, SSM only creates a multicast distribution tree to a specific multicast source address. SSM is considered more secure than non-source-specific multicast, because the multicast client must know both the multicast destination address and the multicast source address in order to join the multicast group. To create a source-specific multicast tree, SSM relies on IGMPv3 signaling from multicast hosts. IGMPv3 includes the multicast source address in the multicast join request. Because current-generation STBs do not support IGMPv3 signaling, the AR can be configured to map IGMPv2 requests received from the aggregation network to PIM SSM (S, G) (source, group) messages in the distribution network. This translation process maps the multicast destination address specified by the STBs in IGMP messages to a combination of multicast source and destination addresses in PIM messages. Release 1.1 of the solution uses SSM mapping at the AR to provide SSM support for STBs that do not support IGMPv3.
Benefit: Anycast Support
When the AR is configured as the Layer 3 edge for video, the distribution network can take advantage of "anycast" support for either the load balancing or the fast failover of video encoders. IP multicast technology natively supports the ability for "anycasting" of IP multicast sources. With anycasting, one configures two or more multicast sources that are sending to the same IP multicast group (with the same multicast destination address) and have the same IP source address. When used with PIM sparse mode, IP multicast technology uses a reverse path lookup to determine which IP source is closest to any particular PIM edge node. The result is that the replication path for a single multicast group can consist of a separate multicast tree for each broadcast encoder. Figure 3-5 illustrates the use of anycasting for load sharing between multiple video encoders.
Figure 3-5 Anycast-Based Load Sharing between Video Encoders
Note that the ability to instantiate multiple multicast replication trees for the same multicast destination is not possible when Layer 2 switching is used. Because each node in a Layer 2 network simply uses IGMP snooping to determine when to replicate packets, anycasting in a Layer 2 domain would result in having the stream from each multicast source replicated to all multicast destinations. Because of this, anycasting is applicable only within the context of a Layer 3 switching environment.
Benefit: Fast Failover of Video Encoders
In addition to supporting load sharing among multicast sources, anycasting can be used to support the fast failover of video encoders. When anycasting technology is combined with the ability of the network to detect the failure of an encoder, routing protocols reconverge. This reconvergence results in the reverse path from the ARs to the DER being recalculated to take into account that the location of the multicast source that has been changed. The IP reconvergence then triggers PIM to resend a join request along the path to the new multicast source. Figure 3-6 illustrates the use of anycast technology to implement the fast failover of redundant video broadcast sources. Release 1.1 of the solution uses this technology to implement fast failover between redundant broadcast encoders.
Figure 3-6 Fast Multicast Source Failover Using Anycast
Benefit: Asymmetric Networking
Finally, when the AR is configured as the Layer 3 edge for video, the distribution network can be configured to support asymmetric bandwidth for video services in the distribution network. The traffic pattern associated with broadcast video and VoD services is extremely asymmetric. Each video channel or session requires multiple megabits of bandwidth in the downstream direction, while the upstream traffic is limited to control signaling for the service. Asymmetric networking allows the network to be configured for more bandwidth in the downstream direction than in the upstream direction. This reduces the cost of the transport network, because it allows the network provider to take advantage of optical components such as wavelength division multiplexing (WDM) transponders and other optical equipment that can be deployed in a unidirectional manner.
Video Forwarding Architecture
Once the choice is made to position the Layer 3 edge for video at the AR, the aggregation/distribution forwarding configuration becomes fairly straightforward. The service mapping architecture used in the solution results in the GE link from each DSLAM being configured for three separate 802.1q VLANs to aggregate Internet access, voice, and video services. Figure 3-7 illustrates the overall video forwarding architecture discussed in this section.
Figure 3-7 Video Forwarding Architecture
AR Configuration
The AR has a set of interfaces connecting to the distribution network and a set of aggregation interfaces connecting to DSLAMs. The AR is configured to switch packets between the distribution and aggregation interfaces at Layer 3. Separate VLANs are configured for each service on the downstream interfaces connected to the DSLAMs. The video configuration of the upstream interface depends on whether the Layer 2 backhaul technology used for Internet access is based on native Ethernet or EoMPLS aggregation. (For details on native Ethernet and EoMPLS configurations for Internet access, see Internet Access Forwarding.)
When native Ethernet aggregation is used for Internet access service, each upstream port connected to the distribution network is configured to use 802.1q encapsulation (VLAN trunking) and includes separate VLANs for transport services (Internet access) and managed application services (voice and video). The video VLAN of each upstream and downstream port is terminated in a separate Layer 3 switched virtual interface (SVI). This configuration causes video coming in on any physical port to be switched at Layer 3 to any other physical port.
Note
When EoMPLS technology is used for Layer 2 backhaul of the Internet access service, the MPLS tags associated with the EoMPLS tunnels can be used to distinguish between the Internet access service and managed application services such as voice and video. This can be used to simplify the configuration of the distribution network, by (1) configuring a single IP interface for each physical port in the distribution network, and (2) restricting MPLS label distribution only to routes associated with the EoMPLS tunnel endpoints (which are configured as loopback interfaces on the aggregation and distribution edge routers). In this configuration, each distribution port is configured as a Layer 3 routed port on which MPLS tag encapsulation is enabled. Because MPLS label distribution is restricted only to routes pointing to the EoMPLS tunnel endpoints, all traffic associated with managed services such as voice and video remains IP encapsulated.
In the downstream direction, the AR terminates the video VLAN of each GE link from each connected DSLAM in a separate SVI. Each SVI is configured as IP unnumbered, so each SVI obtains its IP address from a loopback interface. Because all of the SVIs obtain their IP addresses from a common configured loopback interface, all of the video SVIs associated with downstream ports can share the same IP subnet. This makes dynamic IP address assignment for the STBs that are aggregated by the AR simpler and more efficient, because the STBs can all share the same IP address pool in a dynamic address server (such as a DHCP server).
DER Configuration
The downstream ports of the DER are configured identically to the upstream ports of the AR. (Refer to Figure 3-7.) Depending on whether Layer 2 Ethernet or EoMPLS aggregation is used for the Internet access service, the video stream is terminated, respectively, in either an SVI bound to a Layer 2 VLAN, or a Layer 3 routed interface bound to the physical port. (For details of the configuration of the upstream ports of the AR, see AR Configuration.)
Note
In the solution, video components such as video servers and real-time encoders are connected to redundant DERs. A load-sharing scheme provides video redundancy. This means that the VoD servers and broadcast video encoders connected to each of these routers are actively sending video during normal operation. Ports connecting video components to a DER may be configured at either physical Layer 3 switched ports or Layer 2 ports terminated in an SVI. To simplify address management, ports connecting VoD servers and real-time encoders may all be configured to be in the same Layer 2 VLAN, which is terminated in a single SVI.
IP Routing
To enable dynamic routing specific to video, a routing process is configured on the ARs and DERs. This routing process is configured only on the video SVI interfaces. This enables the video topology to converge at Layer 3 independently of the topologies for the voice and Internet access services.
Note
Layer 2 Aggregation Alternatives
While the AR may be directly connected to the GE uplinks of the DSLAMs it aggregates, there may be network topologies with insufficient subscriber density to warrant having DSLAMs directly connected to an aggregation router. In these types of topologies, there may be a Layer 2 aggregation network between the DSLAM and the AR.
Note
The solution transport architecture specifies that the AR is where the Layer 3 edge for video should be. The transport architecture also specifies that the AR is defined as the first node in the physical topology that aggregates enough subscribers to require either path or node redundancy for video services. Given these transport requirements, it is important that the Layer 2 aggregation network between DSLAMs and the AR does not include either path or node redundancy. One way to identify such an aggregation network is that it does not require spanning tree algorithms to be configured in order to avoid bridging loops.
When a Layer 2 aggregation network is used between DSLAMs and the AR, it is also important that the number of subscribers aggregated at a single AR not cause forwarding table or ARP table scalability issues for the AR. (For some of the issues associated with forwarding and ARP table scalability, see Issue: Scaling for the Layer 2 MAC Table and Layer 3 Forwarding Table.) A general rule that can be used in network design to avoid scalability issues in the AR is that no more than 30,000 subscribers should be aggregated in a single AR.
The Layer 2 aggregation design described in this section prevents security issues in the DSL aggregation network that are associated with the flooding used in standard bridge-learning algorithms. To simplify the requirements of the aggregation switches, this design assumes that the switches support only standard bridge-learning algorithms, and do not support controlled flooding algorithms that prevent upstream packets from being flooded on down stream links. This design also assumes that aggregation switches are capable of segregating MAC broadcast domains through 802.1q VLAN tagging.
Under the above design assumptions, the Layer 2 aggregation design uses a separate VLAN ID per service per DSLAM. The use of a separate VLAN ID per service per DSLAM means that all MAC layer flooding on the aggregation switch is constrained to a single DSLAM per service. This prevents the security issues associated with MAC layer flooding, but it also means that separate copies of video broadcast channels must be sent to each VLAN—resulting in bandwidth being wasted on the link between the aggregation switch and aggregation router. To prevent multiple copies of video being sent on the link between the aggregation switch and the aggregation router, the Layer 2 aggregation design uses a separate multicast VLAN on which all multicast video traffic is sent. The multicast VLAN carries all broadcast video traffic between the aggregation router and the aggregation switch. The use of a separate multicast VLAN means that the aggregation switch that supports Layer 2 aggregation must be capable of performing IGMP snooping and replication between the single upstream multicast VLAN and the video VLAN on each downstream link. Cisco switches support a feature called Multicast VLAN Registration (MVR) to implement this function.
When the aggregation router is configured to use a Layer 2 aggregation network, the multi-SVI configuration described in AR Configuration, for the downstream aggregation links must be used. In addition to this SVI configuration, the AR must have one additional SVI configured for the multicast VLAN. This VLAN has the IP multicast features described in Multicast Configuration Options, configured on it.
Figure 3-8 illustrates aggregation at Layer 2.
Figure 3-8 Layer 2 Aggregation
Multicast
This section presents the following topics related to multicast:
•
•
•
Overview
A major component of the transport architecture is the multicast transport architecture for video. As stated previously, a Layer 3 forwarding architecture for video is used between the DER and the AR. The video topology is separated from the voice and Internet access topologies by means of a separate VLAN for video. This VLAN carries both unicast VoD streams as well as multicast broadcast-video streams.
PIM for multicast is enabled on the video VLAN interfaces on the DERs and ARs, along with OSPF. This enables a video-specific multicast topology to be built. PIM sparse mode is used for the broadcast video service.
The IGMP/PIM boundary for multicast occurs at the SVIs on the AR that are associated with the GE ports from the DSLAMs. IGMP joins are translated to PIM joins at the SVI.
Source Specific Multicast (SSM) is used in the Layer 3 network. SSM simplifies the operational complexity of configuring a multicast network, because it does not require the configuration of a rendezvous point (RP) to allow multicast forwarding as non-source-specific multicast technologies do. In addition, SSM only creates a multicast distribution tree to a specific multicast source address. SSM is considered more secure than non-source-specific multicast, because the multicast client must know not only the multicast destination address, but also the multicast source address, in order to join the multicast group.
Because SSM builds multicast replication trees that are specific to the IP address of the multicast source, there is an implicit requirement that all multicast join requests (IGMP/PIM joins) must include the address (or addresses) of the multicast source (or sources) in the request. While video STB applications could learn both the multicast source and destination address for each broadcast video channel through the electronic program guide (EPG), current-generation applications receive only the multicast destination address from the EPG. As a result, these applications send IGMPv2 join requests that contain only the destination multicast address in the request. The solution works around this by translating IGMPv2 requests that contain only the destination multicast address into SSM PIM join requests that contain both the multicast source and destination address at the AR. The ability to map the multicast destination address contained in IGMPv2 requests to a source/destination pair is called SSM mapping. To map between multicast destination addresses and source/destination pairs, SSM mapping can be configured to use either statically configured maps on each AR, or the services of a Domain Name System (DNS) server that contains a single map for all ARs. The solution uses the DNS-based approach to simplify the administration of this map.
Figure 3-9 illustrates the multicast features used in the solution with the aggregation and distribution networks.
Multicast Admission Control
The Release 1.1 architectural design supports the ability to perform a network-based connection admission control (CAC) function for the broadcast video service. In some broadcast video deployments, it may not be reasonable to support the transmission of all of the broadcast channels offered by the video service on the links between the AR and DSLAMs at the same time. For example, a broadcast video service may offer 150 standard-definition channels and 20 channels of high-definition television. If the channels are encoded by means of MPEG-2, the bandwidth required to support the transmission of all channels simultaneously is 862 Mbps. To ensure that there is no congestion in the queue used for the broadcast video service, bandwidth must be reserved on the GE aggregation links to the DSLAMs. This can be done by simply subtracting the bandwidth used for broadcast video from the bandwidth pools used by the application components of the other services (such as voice and VoD) that require guaranteed bandwidth. In the example above, if the amount of bandwidth that was reserved for broadcast video was based on supporting all channels simultaneously, only 138 Mbps of bandwidth would be available for voice and VoD. This is not enough bandwidth to implement a reasonable VoD service.
The amount of bandwidth reserved for broadcast video can by controlled by implementing an admission control function for that service. This can be implemented by limiting the number of broadcast streams that are replicated on a particular link. Because the GE aggregation links between the AR and the DSLAM are typically the most likely links to be oversubscribed, they are the best place to enforce a stream limit. When stream limits are used for broadcast video, there is a probability that an IGMP join sent by the broadcast video client application as a result of a channel-change request will fail. Because IGMP signaling has no acknowledgement associated with it, there is no explicit failure indication associated with a failed IGMP join request. Instead, a failed IGMP join request simply results in the requested MPEG stream not being delivered to the STB. The subscriber sees a blank picture as the result of a failed channel-change request. While this user interface is nonoptimal, it is consistent with what video subscribers currently experience when a broadcast channel is not available for some reason.
When a CAC function is used for broadcast video, it is important that the service provider sets the stream limit high enough that subscribers very seldom experience failures as a result of a channel-change request. This can be done by using statistical analysis methods such as Erlang analysis. The statistical analysis described in Static IP Multicast Joins on the AR, is an example of the type of analysis that can be used to determine what the stream limit should be set to in order to ensure a low blocking factor for a group of broadcast channels.
Figure 3-9 Multicast Forwarding Architecture
In the solution, the AR can enforce a maximum broadcast bandwidth limit limiting the number of IGMP joins on the ranges of multicast addresses associated with broadcast video to a configured maximum on the aggregation links that the router controls. This is done by means of the ip igmp limit command. The mapping of video channels to multicast addresses can be done in such a way that the AR can associate the bandwidth for different classes of video (standard definition, high definition, and so on) with different ranges of multicast addresses. IGMP join limits can then be set for each range of multicast addresses.
Caution
For example, a service provider may choose to exclude some video channels from the video CAC function and instead reserve bandwidth for all of the channels that are excluded from that function. This configuration may be useful for managing popular channels that the service provider wants to ensure are never blocked. These channels can be excluded from the CAC function by simply not associating an IGMP limit with their multicast addresses.
When the ip igmp limit command is configured on an AR, that router can enforce a maximum broadcast-bandwidth limit by limiting the number of IGMP joins on the ranges of multicast addresses associated with broadcast video to a configured maximum on the aggregation links that the router controls. The mapping of video channels to multicast addresses can be done in such a way that the AR can associate the bandwidth for different classes of video (standard definition, high definition, and so on) with different ranges of multicast addresses. IGMP join limits can then be set for each range of multicast addresses. For example, a service provider may choose to exclude some video channels from the video CAC function and instead reserve bandwidth for all of the channels that are excluded from that function. This configuration may be useful for managing popular channels that the service provider wants to ensure are never blocked. These channels can be excluded from the CAC function by simply not associating an IGMP limit with their multicast addresses.
Effect of Multicast on Channel-Change Performance
One of the important aspects of a broadcast video service that this solution characterizes is the effect of multicast join and leave latency on channel-change performance. This section documents the multicast configurations that testing has evaluated, and makes recommendations that achieve the following design goals:
•
•
•
Table 3-3 illustrates the major components of channel-change latency. Note that the largest factor in the channel-change delay is the I-frame delay associated with the video decoder. (The I-frame is a keyframe used in MPEG video compression.) As the table indicates, multicast performance should not have a significant effect on channel-change delay.
Table 3-3 Major Components of Channel-Change Latency
Multicast leave for old channel
50
Delay for multicast stream to stop
150 1
Multicast join for new channel
50-300
Jitter buffer fill
200
Conditional access delay2
200-600
I-frame delay
500-1000
1 Assumes that the DSLAM implements IGMP fast-leave processing.
2 The conditional access delay is applicable to broadcast channels that are encrypted by means of a conditional access system (CAS) that modifies decryption keys periodically and carries updated decryption keys in-band in the video stream. The STB must wait for the latest set of decryption keys to be delivered in the video stream before it can perform any decoding. The amount of time associated with this delay depends on how often the CAS sends updated decryption information in the video stream.
Analysis of Multicast Bandwidth vs. Delay
The best approach to use for an IGMP/multicast configuration is based on a tradeoff between bandwidth and delay. IP multicast natively supports the ability to perform replication on a stream only when that stream is requested by a downstream device. While IP multicast and IGMP natively support dynamic replication, each can be configured always to replicate multicast data for a particular channel or channel group to any node in the network. When a channel or channel group is always replicated from the source to a particular node, that node is said to be configured for static joins of the channel or channel group. The benefit of configuring static joins at a particular node is that no channel-change latency is associated with dynamic signaling and replication from the source to the node on which static joins are configured. The down side of configuring static joins at a node is that the video streams for the channels that are statically joined are always sent whether a subscriber is watching them or not.
Statistical analysis can be used to determine when the benefits of static joins (less channel-change latency) outweigh the costs (additional bandwidth usage). This section describes the statistical analysis that was done as part of the solution to determine the recommendations for where in the network static joins should and should not be configured.
The behavior of a population of subscribers can be modeled statistically to determine, for a population of subscribers, the probability of at least one subscriber in the group being tuned to a set of television channels. If the probability of at least one subscriber being tuned to each of the channels in a broadcast channel group is fairly high, then the amount of bandwidth that is saved by performing dynamic joins on that group of channels is statistically insignificant. When statistical analysis shows insignificant bandwidth savings for a group of channels, static joins can be used on those channels without having a significant impact on the amount of bandwidth on the GE aggregation links.
The factors used in this analysis included the following:
•
•
•
The number of video subscribers served by a particular node depends on where that node is located in the network. Based on common expected video service take rates, the number of subscribers served by a DSLAM is typically about 500 while the number of subscribers served by an AR is typically about 5000.
The following is a statistical analysis model that is helpful in determining when to use dynamic joins, and when to use static joins.
Analysis of Dynamic Joins in a Video over IP Environment
Each subscriber is modeled as a random process selecting a channel to watch according to a given probability distribution across all possible channels. Given a group of channels, we would like to calculate the average number of channels in use, given the "popularity" probabilities of the channels. Because we are interested in determining the average number of channels in use, we can consider the channels to be probabilistically independent of each other and consider the channels one at a time.
For a single channel, the probability that this channel is idle is calculated as follows:
Let
p = P{a subscriber tunes to this channel}
N = number of subscribers subtended by the given AR or DSLAM
so that
P{channel is idle} = (1-p)N
For multiple channels, we sum the above expression.
Let
C = number of channels
pk = P{a subscriber tunes to kth channel}
so that the average number of channels in use, CIU, is
Channel-Change Latency Probabilities
When subscribers change channels, if they change to a channel that is not part of a static join and that no one else is watching, they experience some latency while the dynamic join is established before they can view the channel's content.
We assume that when there is a channel-change event, the probability a particular channel is changed to is proportional to that channel's popularity. This assumption can be combined with the above calculated P{channel is idle} and the knowledge of which channels are associated with static or dynamic joins to determine the probability that a given channel change results in the latency associated with establishing a dynamic join.
Let
D = the set of all channels involved in dynamic joins
PL = P{a channel change experiences latency due to a newly created dynamic join}
so that
Analysis
Given a set of channels with probabilities pk, they can be ranked from highest to lowest pk. Then, once they are ranked, we can have a cutoff value so that channels with higher pk get a static join and those with lower pk get a dynamic join. The questions then are, for a given cutoff,
•
•
As an example, consider a 150-channel system with an exponential decay function for
P{subscriber tunes to kth channel}
Figure 3-10 graphs the channel popularity for this example.
Figure 3-10 Channel Popularity for a 150-Channel System
Figure 3-11 add curves showing the probability that a given channel is busy for subscriber bases of 200 (at a DSLAM) or 5000 (at an AR).
Figure 3-11 Probability a Given Channel is Busy for Subscriber Bases of 200 (at a DSLAM) or 50000 (at an AR)
The important thing to note here is that the P{busy} curve shifts dramatically to the right when the number of subscribers is increased from 200 to 5000.
Figure 3-12 and Figure 3-13 show the tradeoff between average bandwidth requirements and channel-change latency probability for a DSLAM and an AR, respectively. The horizontal axis is the fraction of channels moved from a static join to a dynamic join. The two curves show the bandwidth required (as a percentage of bandwidth required in the static join case) and the probability of channel-change latency. (The two curves in each figure are shown on different scales to make them both visible.)
Figure 3-12 Tradeoff Between Average Bandwidth Requirements and Channel-Change Latency Probability for a DSLAM (200 Subscribers)
Figure 3-13 Tradeoff Between Average Bandwidth Requirements and Channel-Change Latency Probability for an AR (5000 Subscribers)
Figure 3-12 shows the tradeoff for the DSLAM (200 subscribers). There seems to be substantial opportunity in using dynamic joins where about half the bandwidth can be saved with a channel-change latency probability of about 1 in 50 (0.02). (See the vertical black line in the graph.)
Figure 3-13 shows the tradeoff for the AR (5000 subscribers). In this case, the best possible bandwidth savings is about 20%, even with all channels in dynamic joins. Here the channel-change latency probability is uniformly low, with a maximum value of about 1 in 300.
From the statistical analysis results described above, you can see that there is a typically a significant bandwidth savings to be gained (~60%) by using dynamic joins at the DSLAM. Because of this, we recommend that the multicast configuration models used on the links between the AR and the DSLAM take advantage of the dynamic replication capabilities native to IP multicast. Also from these results, it can be seen that the benefit of using dynamic vs. static joins at the AR depends heavily on the popularity of a channel or channel group. It may be best to join popular channels statically, and join less-popular channels dynamically. Static IP Multicast Joins on the AR, describes additional analysis that was performed to determine when it is best to perform static vs. dynamic joins of a channel group on the AR.
Multicast Configuration Options
From the above analysis, the solution architecture assumes that multicast traffic is replicated by means of dynamic Internet Group Management Protocol (IGMP) signaling on the GE aggregation links between ARs and DSLAMs, and also on the DSL access links between the DSLAM and HAG. The following sections detail the multicast configuration options included in the solution.
IGMP-Based Replication in the DSLAM
Because the DSLAM performs packet switching at Layer 2, it must use a Layer 2 method of implementing multicast replication based on dynamic signaling. In the transport architecture, the DSLAM performs multicast replication by means of IGMP snooping. A Layer 2 switching node that implements IGMP snooping uses the IGMP state machine to determine when to perform multicast replication to a particular link.
Transparent IGMP Snooping vs. IGMP Proxy-Routing Functionality
Note that the recommendation for the transport architecture is to use transparent IGMP snooping and not an IGMP proxy function. IGMP snooping, as defined in the DSL Forum WT-101 specification, is a function whereby the DSLAM uses IGMP messages and the associated IGMP state machine to determine when to perform replication of an incoming multicast stream on outgoing DSL lines. When transparent IGMP snooping is used, the DSLAM appears totally transparent to the IGMP signaling path. It does not modify IGMP messages in either the upstream or downstream directions. WT-101 defines IGMP proxy-routing as a function whereby the DSLAM acts as an IGMP router to STBs, and as a host to upstream routers. With an IGMP proxy-routing function, the DSLAM can statically join multicast streams coming from the AR and replicate them on demand, based on IGMP messages coming from the STBs.
IGMP proxy-routing functionality is not recommended on the DSLAM for a couple of reasons. First, the IGMP proxy-routing function complicates both the operation and configuration of IGMP signaling. This is because the signaling path is now split into two separate IGMP sessions between the STB and the AR. Second, the main benefit of an IGMP proxy function is to allow the DSLAM to join multicast groups statically from the AR and perform dynamic replication to the DSL line. As shown from the analysis in Analysis of Multicast Bandwidth vs. Delay, the benefits of statically joining broadcast channels at the DSLAM (decreased channel-change latency) are far outweighed by the cost (additional bandwidth on the GE aggregation links).
Note
Cisco has load tested IGMP signaling on the Cisco 7600 series, for example, and no join or leave performance degradation was experienced with over 10,000 IGMP messages (join/leaves) per second. Thus, although the Release 1.1 multicast architecture does not require IGMP report suppression on the DSLAM, using this report suppression feature does not cause any issues with the multicast architecture.IGMP Immediate Leave Processing
To meet the channel-change time requirements, the DSLAM must perform IGMP snooping with immediate leave processing. Immediate leave processing, as defined by WT-101, is a modification of the normal IGMP Version 2 host state machine. In IGMPv2, when a router (IGMP server) receives an IGMP leave request from a host (IGMP client), it must first send an IGMP group-specific query to learn whether other hosts on the same multi-access network are still requesting to receive traffic. If after a specific time no host replies to the query, the router stops forwarding the traffic. This query process is required because, in IGMP Versions 1 and 2, IGMP membership reports are suppressed if the same report has already been sent by another host in the network. Therefore, it is impossible for the router to know reliably how many hosts on a multi-access network are requesting to receive traffic.
The requirement of making IGMP queries and waiting for a response can be removed if there is only a single video STB per DSL line that is making IGMP requests. In this case, when an STB sends an IGMP leave request, the DSLAM can safely and immediately stop sending the multicast stream down the DSL line from which the request came. The ability for a node that supports IGMP snooping to stop sending a multicast stream immediately on the receipt of an IGMPv2 leave request is called immediate leave processing. The solution requires that DSLAMs support IGMP snooping with immediate leave processing.
However, IGMP snooping with immediate leave processing does not work when more than one STB is connected to a DSL line. The problem with immediate leave processing is that if two STBs attached to the same DSL line are tuned to the same channel, the first STB that tunes off that channel causes the DSLAM to stop sending the multicast stream for that channel. This in turn causes the second STB to stop receiving video. The workaround for this problem requires additional functionality in both the STBs and the DSLAM. STBs must always send IGMPv2 join and leave requests during a channel-change operation, independently of whether other STBs on the same network segment are currently joined to the same multicast group. The DSLAM must keep track of the IP source address associated with each IGMP join and leave request. The DSLAM stops sending a multicast stream to a particular DSL line when all of the IGMP hosts (as specified by the IP source address in each IGMP message) have issued IGMP leave requests. (In fact, these modifications to the IGMPv2 state machine are required in order to make IGMP hosts compliant with IGMPv3.)
Multicast Replication in the AR
Because the AR forms the IGMP/PIM boundary, multicast replication is triggered by IGMP messages that are received on the GE uplinks from the DSLAMs.
Because the AR potentially aggregates many subscribers, it must be capable of processing a high volume of IGMP join and leave requests if many subscribers are changing channels at the same time.
The solution testing effort characterized the performance of IGMP on the AR by flooding the AR with a constant rate of IGMP join and leave requests, in order to determine the effect on CPU performance in the AR, as well as on the network multicast join delay that contributes to the channel-change performance experienced by an STB. To determine that, an IGMP host makes an IGMP join request for a multicast address that is currently not being sent on the GE aggregation link while the AR is being flooded with IGMP join and leave requests for a different multicast address. The test measures the amount of time it takes from the time the join is sent until the time the stream is delivered, both when the AR is not busy and when it is under various IGMP load conditions.
The performance test described above will be used to provide a recommendation for the maximum number of video broadcast subscribers that should be aggregated by the aggregation router platforms that will be tested as part of the solution when they are configured to use IGMP snooping. To provide this recommendation, the results of the performance test will be first translated into a maximum number of joins / leaves that the aggregation router can process per second before significantly affecting channel change performance. An increase in IGMP join latency of 500 msec from an unloaded to a loaded condition will be considered a significant increase in this testing.
The maximum IGMP performance can be translated to a maximum number of subscribers by using a worst-case channel-change scenario. The scenario used for this recommendation is an event where a significant percentage of the subscribers viewing a popular channel tune off from that channel at the same time (within 1 second). An example of this type of event is a commercial break during a popular program. The IGMP performance required for this type of event can be determined by taking into account the total number of video subscribers, the popularity of the channel, and the percentage of subscribers that tune off that channel at the same time. The performance requirement can obtained by using the following formula:
IGMP_Perf = Num_Subscribers * Channel_Popularity * TuneOff_Factor
where
IGMP_Perf (IGMP performance) = Number of IGMP joins or leaves that a platform can process in 1 second
Channel_Popularity = Percentage of total subscribers tuned to a popular channel.
TuneOff_Factor = Percentage of subscribers that tune off the channel at the same time. (Solution worst-case assumption is 50%.)
In deployments where performance testing shows that IGMP snooping should not be used by itself on the AR, the AR can be configured to use static IP multicast joins, or the DSLAM can be configured to use IGMP snooping with IGMP report suppression on the DSLAM. Both of these alternatives are described below.
Static IP Multicast Joins on the AR
If the AR is configured to use static IP multicast joins, all of the multicast streams that are configured with static joins are sent through the distribution network to the AR independently of whether or not IGMP requests have been made by STBs.
Statistical analysis can be used to determine when the use of static joins in the AR does not result in a significant amount of additional bandwidth on the GE aggregation links. The results of this statistical analysis are shown below.
Each service provider must decide, for each channel group, whether that channel group should be a static join or a dynamic join, based on a balance of configuration overhead vs. delay probabilities. Table 3-4 summarizes the factors and formulas used in this analysis.
Figure 3-14 illustrates the results of the statistical analysis model for bandwidth/latency tradeoff at the AR. Here fixed values are used for the number of channels in a channel group (C = 50) and the number of subscribers served by the node (N = 5000). Note how bandwidth and latency vary with average channel popularity.
Figure 3-14 Bandwidth and Latency vs. Channel Popularity: 5000 Subscribers at the AR
Based on the above, we can make a general recommendation that channel groups with an average per-channel popularity of 0.05% or less should be joined dynamically at the AR, while channel groups with an average per-channel popularity of greater than 0.05% could be joined statically.
Note
IGMP Snooping with Report Suppression on the DSLAM
In situations where there are issues with the configuration shown in IGMP-Based Replication in the DSLAM, IGMP snooping on the AR can be combined with IGMP snooping and report suppression on the DSLAMs to provide a more scalable IGMP snooping configuration. When IGMP snooping with report suppression is configured on a DSLAM, the DSLAM forwards only the first IGMP join request for a particular multicast address on the upstream GE link. In addition, the DSLAM sends an IGMP leave request only when it sees a single DSL link currently joined to the multicast stream. This behavior reduces the number of IGMP joins and leaves that the AR must process, enabling the AR to scale to a larger number of subscribers.
Note
IGMP Functionality in the STB
As described in Broadcast Client, the broadcast client in the video STB is responsible for implementing channel-change requests from a subscriber by issuing an IGMP leave followed by an IGMP join.
Because the bandwidth on the DSL line is often limited, the broadcast client on the STB typically implements the channel-change function by sending an IGMP leave, waiting for the video stream from the channel that is being tuned away to stop, and then an IGMP join. The broadcast client must support IGMPv2, because version 2 is the first release of IGMP that provides the ability for a client to signal explicitly when it wants to leave a multicast group. Broadcast clients that support IGMPv2 should also send IGMP joins during a channel change, independently of whether other STBs have also sent IGMP joins for the same channel.
Note
Broadcast clients should also support IGMPv3. In addition to IGMP state machine enhancements, the support of IGMPv3 by the broadcast client enables the client to specify one or more IP source addresses of broadcast encoders from which it wishes to receive the broadcast channel. To support this function, the electronic program guide (EPG) must be updated to send both the multicast group address as well as a list of the IP addresses of real-time encoders that may be used for each broadcast channel. When the broadcast client as well as the EPG are updated to support IGMPv3, the multicast solution is significantly simplified, because Source Specific Multicast (SSM) is supported from the STB all the way to the real-time encoder. As a result, there is no need to turn on SSM mapping in the AR.
Internet Access Forwarding
Because different services in the transport architecture use different logical topologies, the forwarding architecture for Internet access may be different from that for video. The Internet access forwarding architecture used in the solution provides an example of how Internet access can be implemented alongside a video service.
The solution uses Layer 2 forwarding in the aggregation and distribution networks for Internet access. An example Internet access service that could be implemented by this type of architecture would be PPPoE aggregation to a broadband remote-access server (BRAS) that is connected to the DER. The solution transport architecture supports both VLAN per service (N:1) and VLAN per subscriber (1:1) models in the aggregation network for Internet access service.
Release 1.1 includes two example Layer 2 forwarding architectures for the Internet access service:
•
•
Both of these models are described in the following sections:
Native Ethernet Aggregation
The native Ethernet aggregation model tested in Release 1.1 uses 1:1 VLANs for Internet access service and N:1 VLANs for voice and video services. As explained in Solution Transport Recommendations Based on WT-101, most current-generation DSLAMs do not support the ability to impose the 802.1ad VLAN tags required for 1:1 VLAN architectures that need to scale to more than 4096 subscribers. Because of this, the 1:1 VLAN model used in Release 1.1 for Internet access is limited to DSLAMs that support dual GE uplinks. The dual GE uplinks enable the DSLAM to map managed application services (video and voice) to one GE uplink, and transport services (Internet access) to the other GE uplink. This mapping enables the DSLAM to encapsulate both the 1:1 Internet access service and the N:1 video and voice services by means of 802.1q encapsulation. Because 1:1 services and N:1 services are segregated on different GE uplinks, the AR can impose an outer S-Tag on packets arriving on the GE link associated with 1:1 services, while not modifying the encapsulation of packets arriving on the GE link associated with N:1 services. Figure 3-15 illustrates the operation of a dual GE uplink DSLAM for 1:1 VLANs when the multi-VC access architecture described in Multi-VLAN Access Architecture, is used, while Figure 3-16 illustrates the operation of a dual GE uplink DSLAM for 1:1 VLANs when the EtherType access architecture is used.
Figure 3-15 Multi-VC Access with Dual GE Uplinks and 1:1 VLANs for Internet Access
Figure 3-16 EtherType Access with Dual GE Uplinks and 1:1 VLANs for Internet Access [fix spelling]
Because the Ethernet aggregation architecture of Release 1.1 uses 1:1 VLANs for Internet access, the S-Tag that is imposed by the AR is different for each DSLAM connected to it. This configuration makes the VLAN topology look like a hub-and-spoke topology with a separate logical network between each DSLAM and the two DERs. Spanning tree is configured to break the link between the DERs to avoid a forwarding loop. After the spanning tree converges, the VLAN topology looks like separate point-to-point connections between each AR and the DERs. This logical topology conserves MAC address forwarding entries on both the ARs and DERs, because each VLAN now connects only two physical ports. MAC learning algorithms are not needed when a logical topology consists of only two physical ports, because each MAC frame that arrives at one port is always sent on the other port. Figure 3-17 illustrates the Layer 2 forwarding model used in the solution for the Internet access service (native Ethernet aggregation).
Note
Figure 3-17 Layer 2 Forwarding for Native Ethernet Aggregation
Internet traffic (1:1)
QinQ—SP outer tag removed
Dot1q tunnel
QinQ—SP outer tag added
•
•
EoMPLS Aggregation
EoMPLS backhaul is sometimes preferred for Layer 2 aggregation over native Ethernet, because MPLS supports traffic engineering functionality and faster reconvergence for link failures than spanning-tree-based algorithms.
The EoMPLS aggregation model used in Release 1.1 uses the MPLS tags associated with the EoMPLS tunnels to distinguish between the Internet access service and managed application services such as voice and video. This can be used to simplify the configuration of the distribution network, by (1) configuring a single IP interface for each physical port in the distribution network, and (2) restricting MPLS label distribution only to routes associated with the EoMPLS tunnel endpoints (which are configured as loopback interfaces on the aggregation and distribution edge routers). In this configuration, each distribution port is configured as a Layer 3 routed port on which MPLS tag encapsulation is enabled. Because MPLS label distribution is restricted only to routes pointing to the EoMPLS tunnel endpoints, all traffic associated with managed services such as voice and video remains IP encapsulated.
As explained in Service Mapping in the Aggregation Network, the use of N:1 VLANs in the aggregation network for Internet access implies that the DSLAM must support a method to provide subscriber line ID (SLID) for transport sessions as part of the session establishment process, Because the solution focuses on video services, service-specific features such as SLID for the Internet access service were not tested as part of the solution.
Figure 3-18 illustrates the EoMPLS forwarding model used in Release 1.1.
Figure 3-18 Forwarding Model for EoMPLS Aggregation
Each DSLAM is configured to provide a unique 802.1q VLAN tag for the Internet access service. This VLAN is injected into an EoMPLS pseudowire at the AR by configuring a Layer 3 interface for the VLAN and connecting the Internet access VLAN to an EoMPLS pseudowire by means of the Cisco IOS xconnect command. This command multiplexes one or more EoMPLS tunnel or pseudowires through an MPLS label switch path (LSP). The EoMPLS tunnel is configured between loopback interfaces on the ARs and DERs. The EoMPLS tunnel is terminated at the DER in a Layer 3 interface configured on a GE port connected to the BRAS. Multiple DSLAMs can be aggregated onto the same upstream port by terminating multiple EoMPLS tunnels on the same Layer 3 interface.
Redundancy to protect against link failures in the Internet access service can be implemented by using the MPLS Traffic Engineering Fast Reroute feature. MPLS supports MPLS fast reroute by configuring diverse LSPs between the two endpoints of the pseudowire. This method provides resiliency combined with 50-msec failover for link failure, but it does not provide resiliency in the case of the failure of a node that terminates the EoMPLS tunnel.
To provide resiliency for node failures such as the failure of a DER or a BRAS, the MPLS Virtual Private LAN Service (VPLS) feature can be used to connect the Internet access VLAN to a set of VPLS tunnels that are terminated in redundant DERs.
Note
For more information about VPLS, see "Virtual Private LAN Services (VPLS)" at the following URL:
http://www.cisco.com/en/US/products/ps6648/products_ios_protocol_option_home.htmlVoice Forwarding
Because the solution transport architecture uses an N:1 VLAN architecture in the aggregation network, the forwarding architecture for voice services may be different from that for video and Internet access. Depending on the VLAN and forwarding configuration in the distribution network, voice traffic may be forwarded as part of the Internet access topology, as part of the video topology, or in a separate topology all by itself.
If voice is forwarded as part of the Internet access topology, it is forwarded at the AR at Layer 2 to the BRAS, along with the Internet access traffic. In this model, the BRAS must provide QoS for the voice service as well as enforce the service level agreement (SLA) for the Internet access service. This model may be used by service providers who have already deployed a voice service through the BRAS and want to use the solution architecture to provide video services through a distributed IP network-based approach.
If voice is forwarded through its own logical topology, it is carried on a separate VLAN in both the aggregation and distribution networks. If a separate VLAN is used for voice, a separate routing process is configured for voice and includes all of the voice interfaces on the AR and DERs.
The voice forwarding architecture tested in the solution provides an example of how a voice service may be implemented alongside video and Internet access services. In Release 1.1 testing, the voice service is forwarded by means of the video topology. This means that voice packets are forwarded through the distribution network on the same logical topology that is used for video services. (See Figure 3-7.)
Management
Aspects of management include the separation of services, the management of address spaces, element and network management systems, and service monitoring. These topics are addressed below:
•
Management Transport
The hybrid architecture supported in Release 1.1 of the solution provides the flexibility to allow service providers either to manage each service independently or use a common infrastructure for all services. The level of sharing among services can be controlled by configuring a subset of the IP address to be shared, and deploying common infrastructure components within the shared IP address space. A service provider could thereby share some components such as DHCP and DNS servers, while making other components such as VoD servers specific to the video service.
Solution testing included the configuration and testing of the scenarios where DNS and DHCP servers are shared across services. In Release 1.1, a separate management subnet is configured for components that may be shared across services such as DHCP and DNS servers as well as management hosts and as element management systems (EMS) and network management systems (NMS). These components are connected to the DER through either a separate physical port or a separate VLAN than are devices associated with video or voice services. Also in Release 1.1, the address spaces associated with different services such as voice and video are separated by configuring a separate routing process per service. The management subnetwork can be shared across services by including the interface associated with that subnetwork in the routing process associated with each service.
DHCP Configuration
To enable dynamic address allocation for the devices in the home, the network is configured to support Dynamic Host Configuration Protocol (DHCP). Because the AR is the Layer 3 edge device, DHCP relay functionality is configured on the downstream video VLAN interface of that router. The helper address used with DHCP relay points to a DHCP server located in the management network.
Release 1.1 supports a segmented address allocation scheme that uses a separate DHCP address pool per service. With segmented address allocation, there may be separate address pools associated with each service terminated in the AR. Each address pool is shared by all of the devices aggregated by the AR that are associated with a particular service.
STB Identification and Authorization
As described in Electronic Program Guide, the EPG component that is part of video middleware is responsible for authenticating subscribers for both the broadcast video and VoD services. Because video subscribers are authenticated at the application layer by video middleware components, there is no requirement for the network to authenticate the video subscriber at the transport layer. While subscriber authentication is not required, service providers may choose to provide a level of authentication for the video STB to ensure that this component is authorized for use on the network. This section identifies some of the DHCP-based methods that may be used for this simplified form of identification.
In some environments, the service provider may choose to identify and authenticate video subscribers for DHCP purposes by identifying the DSL port that connects the subscriber's STB to the network. This information may be used in addition to the MAC address of the subscriber's STB to enable the service provider to make a more stringent check that the STB is authorized for use on the network. In these environments the DSLAM must be capable of snooping DHCP requests from devices in the home network and inserting a DSL port ID in the DHCP request by means of DHCP option 82. The DHCP server can then extract this port ID from the DHCP request and use it to identify the subscriber. DHCP option 82 is described in RFC 3046.
Note that because the AR is acting as a DHCP relay agent in the solution, a DSLAM that supports DHCP option 82 appears as a trusted downstream (closer to the client) network element (bridge) between the relay agent (the AR) and the client (the STB). In this mode the DSLAM inserts DHCP option 82 information but does not set the "giaddr" field in the DHCP request. In addition, because the DSLAM is not acting as a DHCP relay agent, it does not modify the destination MAC address of the DHCP request but simply forwards this address by means of Layer 2 forwarding. The DSL Forum WT-101 specification specifies both the DSLAM requirements for DHCP option 82 as well as a recommended common format for providing DSLAM and line ID as part of the relay-agent information included with DHCP option 82.
EMS/NMS
Release 1.1 of the solution does not include the integration of element management or network management systems into a video transport solution. The Cisco command line interface (CLI) is the method of configuring the Cisco platforms included in the solution.
Redundancy
The solution addresses fast recovery from the failure of video infrastructure components, as well as of network components in the distribution network, such as physical links or network switching components. Solution testing looked at the recovery characteristics associated with failures of video components such as the VoD servers used for on-demand services and the real-time encoders used for broadcast services.
Note
In addition, solution testing has determined how to optimize the network reconvergence time associated with the failures of links in the distribution network, as well as with the failure of a DER.
This section discusses two types of redundancy:
•
Video-Infrastructure Component Redundancy
Figure 3-7 illustrates how the transport architecture supports the redundancy of video infrastructure components such as VoD servers and real-time encoders. The solution test bed included redundant video pumps and real-time encoders attached to redundant DERs.
The solution relies on application-layer failover between the redundant video pumps attached to the DERs in one or more video headends. The video server must support the ability to load-balance VoD sessions between the video pumps attached to the redundant DERs. In addition, a video server must be capable of detecting the failure of a video pump and routing new VoD requests from STBs to still-active video servers in the event of the failure of a video pump.
Solution testing has also characterized the recovery time associated with the failure of real-time encoders by using anycast services. As discussed in Benefit: Fast Failover of Video Encoders, anycast technology can be used to support the ability to detect and recover from the failure of a real-time encoder in the time it takes for the network to reconverge. Release 1.1 testing used redundant real-time encoders configured with the same IP source address attached to the redundant DERs to implement the failover of encoders by using anycast.
Testing simulated the signaling of an encoder (broadcast source) failure, which effectively removes the host route for the failed encoder from the DER. The multicast network between the DERs and the ARs then reconverge. The result is that all that the IP multicast trees for the affected broadcast channel consist of sources from the encoder that is still available.
Network Redundancy
The transport architecture uses dynamic IP routing in the distribution network. This means that the failure of either a physical link or a DER should cause both unicast and multicast routing in the IP transport network to reconverge.
Solution testing has characterized the average and maximum reconvergence times for both unicast and multicast in the event of a link failure or the failure of a DER in the distribution network. The reconvergence trigger events that have been characterized by testing include the following:
•
•
•
Average and worst-case reconvergence times were measured by measuring how long video streams are disrupted at the STB. Testing has also characterized the effect on video quality of the loss of IP video to the STB. During testing, the IP video stream was disrupted for different periods of time (50, 100, 200, 500, and 1000 msec) in order to determine quantitatively the effect of this on video quality. Using this reconvergence and video quality information, the service provider should be able to determine accurately the effect of various network outages in various locations in the video transport network.
Solution testing has also determined the optimal configuration for IP unicast and multicast parameters to optimize reconvergence time for video. Finally, testing has determined the ability of the Quality of Service configuration described in QoS Architecture to enable the service provider to degrade on-demand services without affecting video broadcast services in the event of a failure in the distribution network.
Note
Release 1.1 Configurations
Two physical distribution-network topologies based on the transport architecture described in Aggregation and Distribution Transport Architecture were tested for Release 1.1. Both distribution topologies are based on GE rings between the video headend office and video switching offices.
This section presents the following topics:
•
•
•
Overview
One topology (referred to as Configuration 1) uses a 10-GE ring between the VHO and VSOs. This configuration uses symmetric bandwidth around the ring to provide physical link redundancy for all services. The other topology (referred to as Configuration 2) uses 1-GE and 10-GE links between the distribution edge routers and aggregation routers. This topology provides physical link redundancy for the Internet access, voice, and broadcast video services, but it does not provide full redundancy for VoD services. In the event of a link failure, VoD services are degraded without affecting any of the other services through the use of the QoS architecture described in QoS Architecture.
Transport Components
Table 3-5 lists the Cisco transport components tested for both configurations.
Table 3-5 Transport Components Tested for Both Configurations
DER
Cisco router/switch
7600 series, 6500 series
Supervisor
WS-SUP720-3BXL1
N/A
DER <-> AR
10 GE x 4 optic
WS-X6704-10GE
XENPAK-10GB-LR
DER <-> VoD servers
1 GE x 24 optic
WS-X6724-SFP
1GE-SR/LR/DWDM
48-port copper Ethernet
WS-X6748-SFP
1GE-COPP-SR or -LR
AR
Cisco router/switch
7609, 6509
Supervisor
WS-SUP720-3BXL
N/A
DER <-> AR,
AR <-> AR10 GE x 4 optic
WS-X6704-10GE
XENPAK-10GB-LR
AR <-> DSLAM
1 GE x 24 optic
WS-X6724-SFP
1GE-SR/LR/DWDM
1 All line cards use the Distributed Forwarding Card, WS-F6700-DFC3BXL.
Configuration 1: 10-GE Layer 3 Ring
Figure 3-19 illustrates the 10-GE-based symmetric ring topology used in Release 1,1. This topology uses the aggregation/distribution transport architecture described in Aggregation and Distribution Transport Architecture, with the Layer 3 edge for video and voice services at the AR. The Internet access aggregation model tested in this topology is EoMPLS, as discussed in EoMPLS Aggregation.
The 10-GE topology shown in Figure 3-19 provides fiber redundancy for all services. A link cut anywhere on the ring results in traffic from all services being rerouted in the other direction around the ring. Solution testing included a test scenario where the failure of a link in the 10-GE ring and the resulting rerouting of video traffic results in steady-state congestion of video traffic on the remaining 10-GE links. In this scenario the QoS configuration described in QoS Architecture, is used to cause only the VoD flows to be affected while not affecting the broadcast video service at all.
Figure 3-19 Configuration 1: 10-GE Ring
The 10-GE topology shown in Figure 3-19 provides fiber redundancy for all services. A link cut anywhere on the ring results in traffic from all services being rerouted in the other direction around the ring. Solution testing includes a test scenario where the failure of a link in the 10-GE ring and the resulting rerouting of video traffic results in steady-state congestion of video traffic on the remaining 10-GE links. In this scenario, the QoS configuration described in QoS Architecture, causes the VoD flows to be affected, while not affecting the broadcast video service at all.
Configuration 2: 1-GE plus 10-GE Hub and Spoke
Figure 3-20 illustrates a topology that includes 1-GE and 10-GE links between the DERs and ARs. Redundancy is obtained by connecting each AR to a pair of DERs through 1-GE or 10-GE links. This topology uses the aggregation/distribution transport architecture described in Aggregation and Distribution Transport Architecture, with the Layer 3 edge for video and voice services at the AR. The Internet access aggregation model tested in this topology is native Ethernet aggregation, as described in Native Ethernet Aggregation.
Figure 3-20 Configuration 2: 1-GE plus 10-GE Hub and Spoke
Note
http://www.cisco.com/univercd/cc/td/doc/solution/vobbsols/vob1/vbdig/vbdsgn1.htm#wp1167435
Edge Transport Architecture
The edge transport architecture specifies how traffic from the voice, Internet access, and video services are aggregated in separate logical topologies in the aggregation and access networks. The edge network consists of the GE aggregation links between the ARs, the DSLAMs, the DSL links, and the RG.
Note
This section presents the following topics:
•
Overview
Figure 3-21 illustrates the nodes and links in the edge network.
Figure 3-21 Edge Transport Network
As described previously in AR Configuration, voice, video, and Internet access services are separated on the aggregation GE links by assigning each service to a separate 802.1q VLAN. On the DSL link, services are separated by using a separate ATM PVC or a separate 802.1q VLAN tag per service. The DSLAM and RG each include the service with which a packet is associated as part of the algorithms for switching packets.
DSLAM Functions
The solution uses an Ethernet DSLAM that performs MAC layer switching between the ATM VCs on the DSL links and the GE uplink. Mac layer bridging in the DSLAM is enabled through the use of RFC 2684 bridged encapsulation on each VC of the DSL link.
To enable service mapping, the solution architecture supports a number of models for mapping different service topologies to identifiers in the access and aggregation infrastructures. WT-101 Service Mapping, describes the models specified in DSL Forum document WT-101 for mapping services to topology identifiers in both the access and aggregation infrastructures. (The WT-101 specification specifies requirements for IGMP-based replication in DSLAMs.) The solution architecture can support any of the models outlined in that section, as long as the DSLAM implements the features implied by these models. The WT-101 specification specifies the DSLAM tagging and forwarding requirements for each of models outlined in WT-101 Service Mapping.
To support broadcast video services, the DSLAM must support an IGMP snooping algorithm. The DSLAM uses IGMP snooping to determine the DSL ports to which it should replicate incoming IP multicast packets. IGMPv2 snooping allows the DSLAM to track IGMPv2 state for each DSL port and IP multicast destination; it replicates multicast packets to a particular multicast destination to the DSL ports that are actively joined to that multicast destination.
RG Functions
The residential gateway, or RG,1 performs physical adaptation as well as Layer 2 bridging between one or more physical media in the home and the upstream DSL link that uses RFC 2684 bridged encapsulation. The transport architecture does not make any assumptions regarding the physical media used for triple-play services within the home.
The transport architecture also assumes that the home devices that terminate the IP streams for video and Internet access services are typically not integrated into the HAG. Because of this, the architecture assumes that the physical media within the home are capable of transporting IP packets, and use a Layer 2 encapsulation method that can be translated to an 802.3 transport header in a straightforward manner.
For the voice service, the RG may include an integrated voice gateway that translates VoIP into one or more FXS ports that connect to telephone wiring in the home via one or more RJ-11 ports. In this case, there is no need to carry VoIP traffic within the home network.
Table 3-6 specifies some potential physical media and the associated Layer 2 encapsulations that a RG may have to translate to the upstream DSL link with RFC 2684 bridged encapsulation.
In Release 1.1 the RG is responsible for identifying the service topology with which each device in the home network should be associated. This is a very important aspect of the service mapping architecture, because it has implications for how services are delivered to the home network, and on the design of the home network itself.
In Release 1.1, service mapping in the home is implemented by mapping each device in the home to a primary service topology (voice, video, or Internet access). Service Mapping in the Access Network, describes the technologies used to implement service mapping on the DSL line.
The methods by which packets from the home network are mapped to an upstream service topology depend very much on the technology used to the identify the topology on the DSL line itself. The sections below describe how service mapping can be implemented in both Layer 2 HAGs and HAGs capable of Layer 3 Network Address Translation (NAT) for each of methods described in Service Mapping in the Access Network, for mapping service topologies in the access network.
EtherType Service Mapping in a Layer 2 RG
When EtherType-based service mapping is used with a Layer 2 capable HAG, the transport session initiated by each device in the home determines the upstream service topology with which that device is associated. This model assumes that PCs associated with the Internet access service use PPPoE encapsulation for their transport sessions, while STBs associated with the video service use IP encapsulation for their transport sessions. IP phones or embedded VoIP hosts may be associated with either the video or data topologies, and so may use either a PPPoE or IP transport client.
Because service mapping with the EtherType model is essentially performed by the transport host in each home device, the RG itself does not play a role in the service mapping function. The RG simply performs a MAC-layer bridging function between the ports connected to the home network and the upstream ATM VC associated with the DSL line. Figure 3-22 illustrates this mapping topology.
Figure 3-22 EtherType Service Mapping in a Layer 2 RG
Physical Port-Based Traffic Mapping for the Multi-VC and VLAN Access Models
The physical port-based method of traffic mapping takes advantage of the fact that most existing home wiring uses separate physical media for each of the triple-play services. The physical wiring of most homes today includes telephone wiring to telephones for telephony services, coax wiring to television sets or STBs for video services, and may include Category 5 wiring for Internet access services.
An RG could take advantage of this existing wiring to provide service mapping by including an embedded VoIP host for the voice service, terminating MoCA or HPNAv3 over coax wiring for the video service, and terminating either 802.11 or 802.3 for the Internet access service. Because each of these services is terminated in the RG by means of different physical media, the RG can determine the upstream VLAN or ATM VC with which to associate each packet by determining which physical port the packet arrived on. Figure 3-23 illustrates physical port-based traffic mapping in the RG for the multi-VC access model, while Figure 3-24 illustrates the same for the multi-VLAN access model.
Figure 3-23 Physical Port-Based Traffic Mapping with Multi-VC Access Architecture
Figure 3-24 Physical Port-Based Traffic Mapping with Multi-VLAN Access Architecture
One issue with physical port-based traffic mapping is that it enforces a rather rigid mapping of home devices to services. As home network technologies become more ubiquitous, it may not be practical to use the physical media to which a device is connected to associate that device with a particular service. In these environments, it may be more practical to use other methods of mapping between the home network and multiple upstream service topologies. NAT/Layer 3 Functionality, below, provides an example of how to achieve this by using NAT/Layer 3 functionality in a HAG.
NAT/Layer 3 Functionality
To limit the number of IP addresses the service provider must allocate in the home network, and to allow home devices associated with different services to communicate with each other, the RG may include NAT/Layer 3 functionality.
An RG that implements Layer 3/NAT functionality as well as a service mapping function must be capable of mapping each device in a home network to an external transport client that is associated with that service topology. For example, if the voice, video, and data service are each associated with their own service topology, the NAT-capable RG must support a separate DHCP or PPPoE transport client state machine for each of these upstream topologies. Each client may be in a separate external subnetwork and have its own IP address assigned to it.
A Layer 3 RG that supports service mapping must map the source IP address of each device in the home to the IP address of the external client with which the device has been associated as part of the NAT translation function. The mapping between home devices and the associated external client could be statically configured, or it could be learned dynamically through DHCP options such as DHCP option 60 (vendor specific attributes).
A Layer 3 RG implements a local DHCP server that is used to allocate local IP addresses to devices within the home network. All devices within the home network are assigned to a single IP subnetwork whose address is configured in the HAG. When the RG receives a DHCP request from a home device, it creates a NAT entry that maps between that client's IP address and the external address of the client with which the device is associated. This external address can be used to determine the external client and resulting service topology with which the home device is associated.
Because all of the addresses in the home network are assigned by the RG to be in a single IP subnet, devices in the home that are associated with different services can communicate by means of standard IP host functionality. An RG that implements NAT/Layer 3 functionality should also implement standard MAC-layer learning and forwarding functionality between the physical ports attached to the home network. This functionality enables all devices in the home network to communicate independently of the physical port to which they are attached. Because the home network is typically capable of carrying much more bandwidth than is generated as part of the broadcast video service, the MAC-layer forwarding function in the RG may broadcast Ethernet frames received from the DSL port with a destination MAC address in the multicast address range to all downstream ports.
An RG that implements NAT/Layer 3 functionality must implement an IGMP snooping function, in order to enable multicast streams to bypass the NAT logic and be sent directly to the home network without performing address translation. The IGMP snooping function on the RG must not repress any IGMP report messages from home devices. This is needed to enable the DSLAM to implement a fast-leave algorithm that tracks IGMP requests from each home network device. With the introduction of high-definition streams, the home network may not be capable of carrying much more bandwidth than is generated by the broadcast video service. Because of this, the MAC layer forwarding function in the RG should only send Ethernet frames received from the DSL port with a destination MAC/IP address to home network ports in an active IGMP session.
An RG that implements NAT/Layer 3 functionality must support a local ARP function for devices on the home network. The ARP function responds to ARP requests for nonlocal IP addresses (IP addresses that have not been locally allocated by the HAG's DHCP server) with its own MAC address.
An RG that implements NAT functionality must implement stateful inspection for Real Time Streaming Protocol (RTSP, RFC 2326) and Session Initiation Protocol (SIP, RFC 3261) as part of the NAT function. RTSP is the most common session-signaling protocol for a VoD session initiated by video STBs, while SIP is the most common session-signaling protocol for IP telephony applications. Stateful inspection is required by NAT functions that support RTSP and SIP because these protocols specify the IP address and Layer 4 port values for video and voice media streams in the payload of signaling messages.
The following figures illustrate how Layer 3 functionality in the RG can be used to implement a service mapping function on the DSL line. Figure 3-25 illustrates service mapping with the multi-VC access architecture, while Figure 3-26 illustrates service mapping with the EtherType access architecture.
Figure 3-25 NAT/Layer 3 Functionality in the HAG: Multi-VC Access Architecture
Figure 3-26 NAT/Layer 3 Functionality in the HAG: EtherType Access Architecture
QoS Architecture
The Quality of Service (QoS) architecture in the solution is based on the IETF Differentiated Services (DiffServ) Architecture described in RFC 2475. The DiffServ architecture assumes that each node in a transport network that is connected to physical links where congestion can occur must be capable of scheduling packets from different services separately. In an environment where all services are not aggregated at the BRAS, the DSL access links between home access gateways (HAGs) and DSLAMs, as well as the aggregation links between DSLAMs and aggregation routers, can become congested. This means that aggregation routers, DSLAMs, and HAGs must be capable of basic DiffServ functionality.
This section presents the following topics:
•
•
•
Overview of DiffServ Architecture
The DiffServ architecture specifies different requirements for nodes at administrative boundaries than for nodes in the interior of a DiffServ domain. A DiffServ domain is defined as an area where all nodes are configured with the same DiffServ policies for QoS. The edge of a DiffServ domain is the administrative boundary of that domain.
In the DiffServ architecture, nodes at administrative boundaries must implement a superset of the functionality that nodes in the interior of a DiffServ domain implement. Nodes at administrative boundaries must be capable of rate-limiting traffic coming into a DiffServ domain by using a rate-limiting technology such as policing or shaping. Nodes at administrative boundaries must also be capable of marking traffic that is supposed to have different per-hop behaviors, by using separate DSCP code points. An example of a node at the edge of a DiffServ domain in a residential triple-play architecture is the HAG. While the RG is managed by the service provider, the home network typically is not. Because of this, the RG must associate packets that arrive from ports attached to the home network with a service and its associated QoS. The functionality that the RG implements to classify and mark traffic from the home network is an example of the DiffServ functionality required at administrative boundaries.
All nodes in a DiffServ domain that may experience packet congestion must be capable of classifying packets by means of a DiffServ code point (DSCP) and implementing the specified DiffServ per-hop behavior (PHB) accordingly. This functionality must be implemented on nodes in the interior of a DiffServ network as well as on nodes at an administrative boundary.
The DiffServ architecture described in RFC 2475 assumes that all nodes where congestion can occur are capable of implementing QoS functionality at the IP layer. One can extend this basic architecture to nodes that implement QoS functionality at Layer 2 by mapping the DiffServ PHBs specified by DiffServ code points to Layer 2 functionality at the edges of the Layer 2 network. An example of a Layer 2 technology that implements QoS is ATM. The ATM specification defines its own methods of obtaining QoS by using functionality that is part of ATM switching. The ATM traffic-management specification defines classes of service that must be implemented by ATM switching nodes as well as by the nodes at the edge of an ATM network that implement the Segmentation and Reassembly (SAR) function. Examples of services classes defined by the ATM traffic-management specification are Constant Bit Rate (CBR), Variable Bit Rate (VBR), and Unspecified Bit Rate (UBR). In a DSL environment, each of the ATM CoS values can be mapped to a DiffServ PHB without sacrificing the overall QoS requirements of the network.
While the edge of a DiffServ domain represents one level of boundary of trust, service providers (SPs) may choose to implement a second, more secure boundary of trust within the interior of the DiffServ domain. For example, while the functions the RG are considered functions of a boundary of trust, the RG may be compromised because it is not located within the SP's premises. Because of this, an SP may choose to implement additional enforcement functions such as policing at a location in the network that is considered more secure.
DiffServ Architecture in the Solution
This section describes how the Cisco Wireline Video/IPTV Solution uses the DiffServ architecture to implement QoS in support of triple play. Figure 3-27 illustrates the upstream QoS and security functionality used in the solution, while Figure 3-28 illustrates the downstream QoS and security functionality used.
Figure 3-27 Upstream QoS and Security
Figure 3-28 Downstream QoS and Security
Administrative Boundaries
When the DiffServ architecture is mapped to the solution transport architecture, the DiffServ administrative boundaries can be mapped to specific nodes, as discussed below.
In the upstream direction, the administrative boundary is the RG. While the RG is managed by the SP, the home network typically is not. Because of this, the RG must associate packets that arrive from ports attached to the home network with a service and its associated QoS. Because the RG is at the edge of the SP's DiffServ domain, it should be capable of writing a configurable DSCP to each upstream packet, based on the service associated with that packet, by means of the service classification rules described in EtherType Service Mapping in a Layer 2 RG.
The service mapping architecture ensures that the video headend infrastructure resides in a separate logical topology from other services such as Internet access. Because of this, the video headend topology is managed by the SP and can be contained within a single DiffServ administrative domain. If VoD servers and real-time encoders are capable of marking the video streams, as well as the control traffic, with the appropriate DSCP values, then there is no need for the video topology to implement DiffServ edge functionality in the downstream direction. If the VoD servers or real-time encoders are not capable of implementing the DiffServ marking functionality, then the DER should perform this function on their behalf.
DiffServ-to-Layer-2 Mapping
In most DSL/Ethernet aggregation architectures, the access and aggregation networks include nodes that are not capable of supporting IP-layer QoS. In such architectures, the DSLAM and the RG typically implement both packet forwarding and QoS algorithms at Layer 2. QoS in the Access Network, provides the details of how DiffServ-to-Layer-2 Mapping is used to provide proper scheduling behavior in the DSL access network.
Security and Additional Boundaries of Trust
While the edge of a DiffServ domain represents one level of a boundary of trust, SPs may choose to implement a second, more secure boundary of trust within the interior of the DiffServ domain. While the RG limits upstream traffic by means of its ATM-based scheduling functionality, the RG is not located within the SP's premises and may therefore be compromised.
The solution architecture specifies additional functionality that is used to provide additional security for both VoD and broadcast video services, as discussed below.
For VoD services a second upstream policing function is implemented on the aggregation platforms located in either the video switching office or the video headend office. The upstream policing function uses the per-flow policing functionality of the Cisco Catalyst 6000 and Cisco Catalyst 7600 series switches used in the solution to limit the amount of upstream video signaling traffic for VoD to a specified upper limit. With per-flow policing, each upstream flow is recognized dynamically in hardware, and for each new flow a separate hardware-based policer is instantiated. Each policer limits the amount of traffic that is passed upstream to a configured maximum bandwidth and burst size. The bandwidth and burst rate of the upstream policer are determined by the expected maximum bandwidth and burst that video signaling is expected to generate. Any traffic that exceeds the per-flow rate or burst size is dropped.
To limit control-plane-based denial of service (DoS) attacks on the broadcast video service, IGMP access lists can be used on DSLAMs and ARs to restrict multicast join requests to the multicast address range that is known to be valid for the video broadcast service. Any IGMP join requests that fall outside of this address range are dropped. Note that this function is not intended as an enforcement method to limit subscribers to the set of broadcast channels they are authorized to view. The Conditional Access System (CAS) described in Conditional Access System and Encryption Engine is typically used to implement this function.
In addition to IGMP access lists, DSLAMs and ARs can also restrict the rate at which multicast join requests are accepted by the network through the upstream policing of IGMP traffic. In the AR, IGMP policing can be configured by policing all traffic that matches the IP protocol ID for IGMP to a rate that is less than the maximum performance determined by IGMP performance testing.
Triple-Play QoS Analysis
This section describes the analysis behind the DiffServ PHBs and the resulting scheduling QoS configuration recommendations used in Release 1.1 of the solution. This discussion assumes a residential triple-play service with Internet access, voice, and video services. Internet access is assumed to be a best-effort service, with the customer's service-level agreement (SLA) specifying only a maximum (but not a guaranteed minimum) rate. Voice and video are assumed to be managed application services, where the SP provides the subscriber with a video STB and sells the subscriber a video or voice SLA.
An example of a video SLA that an SP may offer is a single channel of VoD or broadcast video delivered to each STB for which the subscriber signs up. The subscriber may sign up for basic or premium-tier broadcast services, and may also sign up for a set of VoD services offered by the provider. The maximum number of STBs that the subscriber may sign up for is limited by the following:
•
•
•
Internet Access
If Internet access is sold as a best-effort service, the DiffServ Default PHB can be used to schedule packets classified as belonging to the Internet access service. The DiffServ Default PHB is described in RFC 2474. The DiffServ PHB provides a best-effort packet-scheduling behavior.
On the aggregation and distribution edge routers, the Default PHB is implemented by using a weighted scheduler that is configured for a minimum bandwidth guarantee. This configuration ensures that Internet access traffic does not significantly affect jitter, latency, or drop for packets associated with the voice or video services.
Voice
End-to-end latency and jitter are very important for a VoIP service. A typical end-to-end jitter requirement for a carrier-class VoIP service is 60 msec. Low jitter and latency are essential to a voice service because the additional delay that results from both factors makes conversations less of an interactive experience, degrading the telephone user's experience.
While delay is an extremely important factor for a successful voice service, the drop requirements for a voice service are not as stringent as they are for video. The reasons that packet drop requirements are not as stringent for a voice service as for a video service are due to digital-to-analog translation algorithms available in current VoIP endpoint implementations. These implementations include a concealment algorithm that can conceal the effects of the loss of a 30-msec voice sample. This means that a packet loss that causes less than a 30-msec loss of digital audio results in an analog signal with no noticeable impairment to the user. With voice concealment algorithms it takes a loss of two or more consecutive 20-msec voice samples to result in a perceptible loss of voice quality. A drop rate of 1% in a voice stream results in a loss that could not be concealed every three minutes when concealment algorithms are taken into account. A 0.25% drop rate results in a loss that could not be concealed once every 53 minutes on average.
Because of this stringent latency requirement, voice services use the DiffServ EF (Express Forwarding) PHB. The DiffServ EF PHB is described in RFC 3246. The EF PHB defines a scheduling behavior that guarantees an upper bound on per-hop jitter that can be caused by packets from non-EF services.
On the aggregation and distribution edge routers, the EF PHB is implemented by means of a priority scheduling algorithm. This algorithm ensures that voice packets can only be delayed by at most one packet serialization time by nonvoice packets per network hop. This delay amounts to a maximum of 12 microsec per hop on 1-GE links configured for a 1500-byte MTU.
Video
While voice has stringent jitter and latency requirements and a relaxed loss requirement, video has a very stringent loss requirement and a relatively relaxed jitter requirement.
Current video-encryption technologies are not resilient to a loss of information in the compressed video stream. As a result, the loss of a single IP packet in a video stream typically causes a noticeable degradation of video quality. The hit to video quality can vary from pixelization across a few frames to a video stream that is frozen for up to 1 second depending on which information in the video stream is lost. The result is that the packet loss requirements for video are extremely stringent. Because of the lack of a concealment algorithm for video, the allowed drop rate for a video service with at most one visible defect per hour is 10-6.
The maximum jitter requirement for video can be determined by examining the maximum channel change delay for broadcast video. Broadcast Video Channel-Change Time, describes the components of channel change delay for a broadcast video over IP service. From Table 3-3, the component of channel-change delay associated with the jitter buffer on the STB is typically around 200 msec; the size of this buffer determines the maximum allowed jitter—200 msec—for a VoIP service.
Because traffic associated with the Internet access service is carried in a best-effort queue, it does not have a significant impact on jitter for video. However, because voice is carried in a priority queue, it does have an impact on jitter for video. The impact of voice on video jitter is minimized by the fact that in most triple-play deployments, the link utilization of voice traffic is not a significant amount of the total link bandwidth. When the relatively loose jitter requirement for video (200 msec) is taken into account, the relatively low link utilization does not result in video jitter above the maximum limit.
Because of the above factors, video flows are scheduled by means of the DiffServ AF PHB. (The DiffServ AF PHB is described in RFC 2597.) While the AF PHB does not provide as stringent a jitter guarantee as the EF PHB used for voice, it can be used to guarantee a maximum jitter/latency and drop rate for a class of packets. RFC 2597 defines four different classes of the AF PHB. To maintain consistency with current IETF DiffServ marking conventions, this document uses the AF4 PHB. This means that all video flows are marked with a DiffServ code point in the AF 4X range. In addition to four scheduling classes, the AF PHB makes it possible to provide four different drop characteristics for each scheduling class. These different drop characteristics are called drop precedence values. Broadcast Video vs. Video on Demand, provides details on how the different availability requirements of broadcast video and VoD traffic can be implemented by using different drop precedence values within the AF PHB.
On the aggregation and distribution routers, the AF PHB is implemented by means of a weighted scheduling algorithm. To ensure the packet loss and drop requirements for video, the weight configured on the video queue should be greater than the combined bandwidth of the traffic associated with both services under normal operating conditions. Broadcast Video vs. Video on Demand, provides specific recommendations for the weight that should be applied to this queue.
Burst Accumulation
The last factor that must be taken into account in determining QoS requirements for video is burst accumulation. Burst accumulation is an instantaneous burst of traffic that is caused by multiple sources of video transmitted through an IP network. If two or more sources are unsynchronized, there is a probability that the packets they generate are transmitted at exactly the same time. If this traffic converges at an intermediate physical link, the link experiences an instantaneous build up of packets. As these bursts of packets are transmitted to downstream routers, the bursts becomes even larger. This is called burst accumulation. Burst accumulation is influenced by a number of factors, including the following:
•
•
•
Burst accumulation may be characterized by means of probability analysis. A typical burst-accumulation analysis shows the probability of a burst of a particular size based on the number of sources, the number of hops between the sources and the receivers, and the amount of video traffic carried on network links. A burst results in either (1) network jitter, if the router has enough buffering for the burst; or (2) a packet drop, if there is not enough buffering to handle the burst.
Figure 3-29 provides an example of what a set of burst-accumulation curves look like. (This is an example only and does not represent the data from an actual simulation.) In this example, the number of hops and the number of video sources have been fixed, with separate curves for different values of video link utilization. Note that as the probability decreases, the maximum delay increases. When probability curves such as these are mapped to network design constraints, the probabilities can be mapped to the maximum allowed end-to-end drop rate for a class of traffic, and the delay can be mapped to the maximum amount of jitter that can be expected for one or more flows within a class of traffic. For video, the maximum jitter number has two implications for the transport network and for the video STB:
•
•
Figure 3-29 Example Burst-Accumulation Curves
When burst accumulation is applied to video, the low allowed drop rate for video (10-6) should be mapped to the same probability value on a burst-accumulation probability curve. From Figure 3-29, this low probability is associated with a relatively high maximum delay. If the probability curves were based on real data, the STBs would need 250 msec of buffering to make up for network jitter to a probability of 10-5. It would also mean that the network would need to provide 250 msec of buffering for these worst-case flows.
In current video deployments, the effects of burst accumulation are dramatically reduced by the fact that there are very few sources of video in the network. In current deployments, the sources of video for VoD services are video servers that are located in video headends. The video servers in each headend stream video only to the STBs served by that headend. In addition, the sources of video for broadcast video services are typically located in a single super headend as well as in local headends. The result is that each STB receives broadcast video and VoD streams from at most two locations in the network.
However, burst accumulation may become more of a factor for video services in the future, as diverse VoD and broadcast video content is distributed to VoD servers and broadcast encoders located at multiple points in the network. When video streams destined to a group of STBs can originate from many different locations in the network, video streams from many different sources may converge at multiple locations in the network. In such an environment, both the jitter buffers on STBs, as well as the amount of buffering available in routers in the network, will need to increase.
Broadcast Video vs. Video on Demand
The QoS requirements for video do not change depending on the service with which the video is associated. For example, the allowed drop rate (10-6) and maximum jitter (200 msec) allowed for both broadcast video and VoD services are the same.
Even though the QoS requirements for these two services are the same, the availability requirements typically are not. As described in Service Availability, the availability requirements for a broadcast video service are typically higher than those of a VoD service. In addition, High Bandwidth, explains why the amount of bandwidth consumed by VoD services is typically much higher than that of broadcast video services in aggregation and distribution networks. Because of the different availability and bandwidth requirements associated with both video services, service providers may decide to reduce the cost of the video transport network by not providing as much backup bandwidth for VoD services as for broadcast video services. When transport networks are designed in this way, a network failure should result in reduced capacity of the VoD service, while not affecting the broadcast video service.
VoD Admission Control, describes the video admission control (VAC) function used in Release 1.1 of the solution. When this functionality is used in the network, a network failure that reduces the amount of transport capacity for video results in the admission control function accepting fewer VoD requests than under normal circumstances. This functionality also results in the failure of existing VoD sessions if the number of existing sessions is greater than the capacity the network can support in its degraded state.
Even with VAC, after a reconvergence event the links carrying video become congested for a period of time if the failover path does not support as much bandwidth as the primary path. This behavior results from a delay—between the time the network reconverges and the time the RSVP state is refreshed— during which the new path for VoD flows is taken into account. QoS can be used to prevent the congestion due to excess VoD flows from affecting the broadcast video service. The drop precedence characteristic of the AF PHB can be used to ensure this behavior in the event of a network failure.
The solution has implemented a VoD priority queueing mechanism that may help users until VAC is available. (For details, see Configuring QoS on DER1.) QoS can be configured to ensure that in the event of a link failure that causes reduced video capacity, only VoD flows are affected. The drop precedence characteristic of the AF PHB can be used to ensure this behavior in the event of a network failure.
Drop precedence associated with the DiffServ AF PHB is implemented in the solution architecture by marking all broadcast video traffic with DSCP value AF41, and marking VoD traffic with DSCP values AF42 and AF43 (high- and low-priority packets, respectively). This marking can be done either by the VoD servers and real-time encoders, or by the service router for VoD servers and encoders that do not support DiffServ marking capabilities.
The two DiffServ drop-precedence behaviors are configured by configuring separate queue thresholds for VoD and broadcast traffic on the weighted queue configured for the AF PHB. Queue thresholds set a limit on the effective queue length for a particular class of traffic that is less than the size of the physical queue. Queue-threshold algorithms are run when packets are received at an egress interface before the packets are entered into the output queue. Threshold algorithms can provide simple algorithms such as tail drop or more complex algorithms such as weighted random early discard (WRED). A tail-drop algorithm compares the current queue length to the length of the threshold configured for a particular class of traffic and drops the packet if the current queue length is greater than the configured threshold. Video packets are not transmitted by means of reliable transport protocols that implement congestion-avoidance mechanisms such as TCP/IP. Consequently, simple threshold algorithms such as tail drop should be used for video, because a WRED algorithm may cause video packets to be dropped even when the configured queue limit is not reached.
To ensure that VoD packets are dropped before broadcast video packets in the event of link congestion, a queue threshold is configured for packets marked with AF42 and AF43. The size of this threshold should be the expected ratio of VoD traffic to all video traffic (VoD + broadcast) on the egress link multiplied by the configured queue length. In the distribution network, the ratio of VoD traffic to all video traffic is often between 50% and 80%. If a link ever gets congested with video traffic because of an unexpected network failure, the queue threshold configured for VoD ensures that VoD packets are dropped before they are entered into the output queue.
Voice plus Video Signaling
Both voice and video services use IP-based signaling to set up and tear down subscriber-initiated sessions. For voice services, Session Initiation Protocol (SIP) is often used to set up and tear down telephone calls between subscribers. For VoD services, Real Time Streaming Protocol (RTSP) is often used to set up sessions to a VoD server that streams on-demand content. The subscriber-initiated signaling protocol used to change channels for broadcast video services is IGMP.
Both voice and video signaling require better than best-effort treatment in order to have an effective service. Drops of signaling packets delay session setup. Since RTSP and SIP normally use TCP as a reliable transport protocol, the additional delay is caused by dropped packets within the period of a TCP retransmission window. The service that is most affected by drops of signaling packets is broadcast video. This is because channel-change latency has the most stringent requirements of the three services described, and IGMP does not include a reliable transport method.
To ensure that voice and video session-setup latency is not adversely affected by interface congestion, voice and video signaling are scheduled by means of a class selector PHB. The recommended DiffServ code point (DSCP) to use for voice and video signaling is CS3. On the AR and DER, the class selector PHB is implemented by means of a weighted scheduling algorithm. To ensure that signaling packets are not dropped in the event of congestion, the weight configured on the queue should be greater than the maximum bandwidth expected for voice and video signaling traffic under normal operating conditions.
QoS in the Aggregation/Distribution Network
In the solution architecture, downstream packet scheduling in the aggregation and distribution networks is implemented by means of line-card-based scheduling algorithms to implement the DiffServ PHBs recommended for the voice, video, and Internet access services. (For discussions of the DiffServ PHB and line-card-based scheduling algorithms used for Internet access, voice, video, and voice and video signaling, respectively, see Internet Access; Voice; Video; and Voice plus Video Signaling.)
Based on the DiffServ DSCP values and associated PHBs for each of the four traffic classes listed above, there is an implied assumption that line cards in the aggregation and distribution networks should support four queues (one priority, three weighted) as well as one threshold to differentiate broadcast video from VoD traffic. Although some of the line cards used in the solution transport architecture support only three queues, this in fact does not turn out to be a problem, because the QoS architecture carries video traffic as well as voice plus video signaling in two weighted queues. Both video and voice plus video signaling use the AF PHB, both classes of traffic can be scheduled by using the same queue on the line card without adversely affecting either class of traffic. Combining both classes in the same queue also simplifies configuration, because queue weights need to be configured for only two weighted queues.
Table 3-7 provides the recommendations for line card configuration using the DiffServ recommendations described in Internet Access; Voice; Video; and Voice plus Video Signaling.
Note
In addition to DiffServ-based scheduling, the aggregation router sets the 802.1p value of packets being sent on aggregation links according to the DSCP value in each packet. Table 3-7 also provides the recommendations for marking the 802.1p value in the Ethernet header based on an IP packet's DSCP value.
If the density of the DSLAM is such that the Ethernet uplink can become congested, the DSLAM must include upstream scheduling functionality. Since Ethernet-capable DSLAMs forward Ethernet frames by means of MAC-layer switching, they typically implement QoS on the Ethernet uplink by using MAC-layer classification techniques. This is done either by (1) associating the incoming ATM VC of an upstream packet with a service and an associated QoS scheduling class, or (2) or by using the 802.1p marking on the packet to associate the packet with a specific scheduling class. DSLAMs compliant with the solution's QoS architecture must be capable of associating the incoming ATM VC of an upstream packet with a service and an associated QoS scheduling class. DSLAMs compliant with the solution's QoS architecture should be capable of using the 802.1p marking on upstream packets, to associate them with a specific scheduling class. Note that this second form of classification on the DSLAM implies that the RG should be capable of marking the 802.1p value of upstream Ethernet frames according to the service with which the RG associates a packet.
1 The downstream queue weight for video is a recommendation that assumes all video traffic consumes no more than 70% of the physical link bandwidth for the link being configured. If the expected ratio of video traffic to total link bandwidth is significantly less, then a lower queue weight may be used.
2 The upstream queue weight for video takes into account only voice plus video signaling, because broadcast video and VoD traffic is unidirectional. The actual value used for the queue weight may vary, depending on the expected ratio of signaling traffic compared to total link bandwidth.
QoS in the Access Network
The RG is located at the DiffServ administrative boundary in the upstream direction. While the RG is managed by the service provider, the home network typically is not. Consequently, the RG must associate packets that arrive from ports attached to the home network with a service and its associated QoS. Because the RG is at the edge of the SP's DiffServ domain, it should be capable of writing a configurable DiffServ code point to each upstream packet according to the service with which it has associated that packet, using the service classification rules described in EtherType Service Mapping in a Layer 2 RG. For HAGs that implement DSCP marking, the DSCP value with which the RG marks each packet based on service classification follows the conventions illustrated in Table 3-8.
The solution provides two potential methods for implementing packet scheduling in the access network. The method used depends on the capabilities of the DSLAM and the HAG. ATM Layer Scheduling, below, describes the required method of packet scheduling based on the ATM layer scheduling methods. MAC/IP Layer Scheduling, describes an additional optional method of packet scheduling, based on MAC/IP-layer scheduling, that DSLAMs and HAGs may use.
ATM Layer Scheduling
Both the DSLAM and the RG include ATM Segmentation and Reassembly (SAR) functionality. The ATM SAR function encapsulates IP packets in ATM AAL-5 frames and segments each frame into ATM cells. The ATM SAR function that is incorporated in HAGs and DSLAMs is typically capable of implementing the cell-scheduling algorithms required for most of the ATM classes of service defined in the ATM traffic management specification. Consequently, the solution QoS architecture requires support for ATM-layer Quality of Service (QoS) for scheduling across the DSL link.
The use of ATM-layer QoS in the RG means that the RG must be capable of mapping the service with which it associated with each upstream packet to the appropriate ATM Class of Service (CoS). The use of ATM-layer QoS in the DSLAM means that the DSLAM must be capable of mapping the incoming VLAN of downstream packets and the service with which that VLAN is associated to the appropriate ATM CoS on the DSL line. In addition, the DSLAM should be capable of mapping the incoming 802.1p value of downstream packets to the appropriate ATM CoS. Note that this second form of classification on the DSLAM relies on the AR to set the 802.1p value in Ethernet frames before they are sent on the aggregation links to the DSLAM.
When ATM scheduling is provided on the DSL line by means of multiple VCs with an ATM SAR function, it provides both scheduling and a link fragmentation and interleaving (LFI) functionality. LFI may be needed for voice services when the amount of upstream bandwidth available on the DSL line is below 400 kbps. LFI is needed in this case because the serialization delay for a single 1500-byte packet could exceed 30 msec on the DSL line (30 msec is about half of the end-to-end jitter budget for a voice service). When multiple ATM VCs are configured on the DSL line, the ATM SAR function breaks each IP packet into a sequence of 53-byte cells and then schedules each cell by means of ATM-based scheduling algorithms. This process ensures that the maximum delay that may be experienced by a voice packet because of the serialization of video or data packets is 1 msec on a 400-kbps DSL link.
Table 3-8 shows the mapping between traffic classes, DiffServ Per Hop Behavior (PHB), and ATM-based scheduling through Class of Service (CoS). Triple-Play QoS Analysis, provides details on the analysis used to determine the mapping used in the solution between services and DiffServ PHBs.
This table also provides a recommendation for configuring the sustained cell rate (SCR) for VBR and CBR virtual circuits (VCs). The expected bandwidth that is used to calculate the SCR for the voice and video VCs can be determined by multiplying the maximum number of voice and video streams by the maximum bandwidth per stream that is expected on the DSL line. Additional bandwidth may also need to be added to take into account voice and video signaling through the VC. Note that the recommended SCR values shown in the table provide about 25% extra bandwidth over what is required to support the expected bandwidth value. This is because the SCR value for CBR and VBR VCs is used to guarantee a minimum rate and also enforce a maximum rate for traffic through that VC. The additional 25% value ensures that the maximum rate enforcement does not degrade the voice and video services during normal operation.
MAC/IP Layer Scheduling
Some DSLAMs and HAGs support the ability to schedule packets by means of DiffServ-based scheduling algorithms. Edge Transport Architecture, describes an optional method that uses 802.1q VLAN tags that the DSLAM and RG can also use to identify the service topology with which each packet on the DSL line is associated. In environments where both the DSLAM and RG support DiffServ-based packet scheduling and 802.1q encapsulations on the DSL line to identify service topology, a single ATM VC can be used between the RG and the DSLAM.
When a single ATM VC is configured between the RG and the DSLAM, the 802.1q marking is used to identify the service topology with which the VC is associated, while either the DSCP value or the 802.1p value is used to classify packets for scheduling. In addition to being able to mark the DSCP value in upstream packets, an RG that supports a single ATM VC configuration should support the ability to mark the 802.1p value in upstream packets on the basis of the service classification described in EtherType Service Mapping in a Layer 2 RG.
Table 3-9 shows the mapping between traffic classes, DiffServ PHBs, and DiffServ-based scheduling algorithms on the DSL line. Triple-Play QoS Analysis, provides details on the analysis used to determine the mapping used in the solution between services and DiffServ PHBs.
1 The downstream queue weight for video is a recommendation that assumes all video traffic consumes no more than 70% of the physical link bandwidth for the link being configured. If the expected ratio of video traffic to total link bandwidth is significantly less, then a lower queue weight may be used.
2 The upstream queue weight for video takes into account only voice plus video signaling, because broadcast video and VoD traffic is unidirectional. The actual value used for the queue weight may vary, depending on the expected ratio of signaling traffic compared to total link bandwidth.
When a single ATM VC is used between the DSLAM and the HAG, IP-based link fragmentation and interleaving functionality may be needed on the RG and DSLAM. LFI may be needed for voice services when the amount of upstream bandwidth available on the DSL line is below 400 kbps. LFI is needed in this case because the serialization delay for a single 1500-byte packet could exceed 30 msec on the DSL line (30 msec is about half of the end-to-end jitter budget for a voice service). DSLAMs and HAGs that support the DiffServ-based scheduling and 802.1q-based service-mapping functionality required in single-VC environments should also support the ability to implement LFI across the DSL line by means of Multilink Point-to-Point Protocol (MLPPP).
VoD Admission Control
Overview
Video on demand can be challenging to keep up with. If the service is more successful than is forecasted, then the bandwidth consumed between the subscribers and the VoD server complex can become oversubscribed. For example, if we assumed a peak concurrency of 20% of all STBs watching on-demand content at the same time, then a central office supporting 10,000 subscribers would need 2,000 streams at peak, or roughly 4 to 7 Gbps, depending on which codec is used to support standard definition. Add high-definition on-demand content, and the bandwidth demand can be quite large.
What happens when demand rises above the peak concurrency that one assumed in designing the network or VoD service capacity? If a particular VoD server complex cannot service a particular request, the request may get rerouted to another VoD server complex that has capacity. If the network capacity over a link or set of links is exceeded, allowing too many VoD sessions to be set up could cause an excessive packet-drop rate for all the video streams, resulting in many subscribers perceiving an outage!
Instead, an integrated network-based admission control can deliver a "could not be serviced at that time" signal to the requesting STB if a video session cannot be supported because of oversubscription anywhere in the network or service. While no one likes a busy signal, the possibility of a mass degradation of the VoD service is much worse! It is easy to imagine that the service could "know" which streams are ending soon, and, thus could provide more-sophisticated busy messages—giving the subscriber more choices for a delayed start of the VoD stream, alternative video service offerings, commerce opportunities, and the like.
Integrated On-Path and Off-Path Admission Control
A VoD admission control (VAC) solution must be able to take into account complex network topologies that have redundant and load-sharing paths in the transport network, as well as access-link utilization or business policies that may be enforcing other types of constraints on the subscriber's service. To do this, the network's routers, in coordination with policy managers and on-demand servers or managers, need to perform a collective admission control function, called Integrated VoD Admission Control (IVAC).
First, Cisco has developed an in-path method to perform admission control for the complex core and distribution network topologies found in next-generation network (NGN) designs. This method uses RSVP for in-path signaling, sent by the VoD server or a component on its behalf before the VoD session is started. The RSVP message traverses the same exact path the VoD session uses, thus tracking in real time any changes in the complex network topologies in the core and distribution layers. Along the path, Cisco IP routers perform a bandwidth accounting function, either allowing the session or denying it if bandwidth is not available for that VoD stream. Having IP, or Layer 3, routing present on every network element, from the VoD server complex to the AR in the central office makes this in-path admission control possible.
Second, to prevent a video stream from being sent to an STB if the access link to a subscriber's home doesn't have enough capacity to carry the stream, the VoD server or a network component in the path mechanism sends a request to an off-path component. This component may be a policy server that is keeping track of the access network, usually a simple and static topology. The policy server can check to see if the access link has enough unused bandwidth, as well as check for business policies that may or may not allow the stream to be supported, in order to allow or deny the session at that time. An off-path component is not recommended as a way to perform admission control for the core and distribution layers, where tracking real-time any changes in the complex network topologies is difficult at best. Consequently, the combination of in-path admission control with an off-path policy server at the edge is the most reliable and efficient way for an admission control solution to decide whether or not a new VoD stream should be allowed to a specific subscriber.
The VAC and related bandwidth accounting, when added to a DiffServ QoS architecture that defines packet marking and scheduling behaviors, can be used to ensure that the video streams allowed to be set up can meet the required 10-6 packet-drop rate. Combining the different admission control technologies with DiffServ QoS allows Cisco to offer a simplified queueing strategy and have all the video traffic shared the same queue. This avoids complex and very costly centralized, hierarchical queueing strategies.
1 Also referred to as a home access gateway, or HAG.
Posted: Mon Jun 26 09:19:29 PDT 2006
All contents are Copyright © 1992--2006 Cisco Systems, Inc. All rights reserved.
Important Notices and Privacy Statement.
