This document provides an overview of the XORP multicast routing architecture. It is intended to provide a starting point for software developers who wish to modify the multicast-related software. The XORP multicast architecture consists of user-level software implementation of multicast routing protocols such as PIM-SM and IGMP. This document provides an overview of the interaction among them, as well as the interaction with the underlying multicast forwarding engine and other parts of the system. As with the other parts of the XORP architecture, the multicast architecture is based on modularity and abstraction. In particular, there is one process per protocol, and each process typically communicates with other processes by using XRLs, the XORP inter-process communication mechanism.
As with the rest of the XORP architecture, the XORP multicast routing architecture is designed such that each of the implemented protocols can be used for both real-world routing and for network simulations. A simulation can be composed of several (UNIX) processes, one per protocol entity controlled by a single process, or all protocol entities running within a single process. For debugging purpose the latter may be simpler to use and control. Further, the all-in-one model might be more scalable. The implication of this is that a protocol implementation must not keep any global state. In routing, the concept of a network port or interface is essential in a sense that a routing protocol entity communicates with the rest of the world through such ports/interfaces (think of this as a protocol entity having several virtual interfaces and each virtual interface corresponds to a physical interface, and those interfaces are used for sending/receiving control messages to/from other routers). In case of PIM for example, the routing entity has several virtual interfaces one per each physical interface it controls. In case of BGP, the routing entity has several ports, one per each BGP peering. In term of C++ implementation, each virtual interface is a class; each protocol entity is a separate class: a single protocol object has several virtual interfaces (see the following figure).
At the protocol level implementation, the virtual interfaces do not posses knowledge about the particular communication method used to send or receive packets. To achieve this, the implementation uses pure virtual C++ class methods (e.g., methods declared such as: virtual int send() = 0;). The declaration of such methods requires that each class must be inherited by a “wrapper” class that implements all pure virtual methods. Then, all the details about the particular mechanism that is used to send or receive packets are in this wrapper class. For example, that class might use XRLs, but it could be any other mechanism that the developer provides (e.g., a third-party simulation testbed environment). Such abstraction makes it very easy to reuse the protocol-specific code for any other purpose, or to create several virtual “nodes” each of them running the particular protocol. Those nodes can be either within the same (UNIX) process or each of them running as a separate process, as long as we have the right mechanism that allows them to communicate with each other.
A physical interface is owned by a single multicast routing protocol, therefore we can have MLD/IGMP running as a library linked with a multicast routing protocol on each of the interfaces owned by that protocol. This solution may not work only if there is more than one protocol that needs multicast group membership information on an interface. The alternative solution is to have a separate MLD/IGMP daemon that takes care of multicast group membership for all interfaces, and then reports that information to all interested parties. Thus, more than one protocol may receive multicast group membership information per interface. Separating MLD/IGMP from the routing code has the advantage of reducing dependency and improving robustness; e.g., if the MLD/IGMP code crashes, the routing protocol can continue running. Further, if we want to upgrade the software, and if we are running MLD/IGMP as a separate process, then the upgrade can be performed by starting a new MLD/IGMP daemon in place of the old one. Therefore, the upgrade does not require to stop the PIM daemon for example, which is very important for routers in real-world operation.
The following figure shows the interactions between the multicast-related modules.
Briefly, those interactions are:
|FEA→MFEA||to propagate vif-related information updates from the FEA to the MFEA.|
|MFEA↔PIM||to send or receive PIM control packets, to forward PIM-related signals from the multicast forwarding engine to PIM, to establish the communication between PIM and the multicast forwarding engine, for the MFEA to join or leave a multicast group on behalf of PIM, etc.|
|MFEA↔MLD/IGMP||to send or receive MLD/IGMP control packets, and to join or leave a multicast group on behalf of MLD/IGMP.|
|MLD/IGMP→PIM||for PIM to monitor multicast membership about local members.|
|RIB→PIM||for PIM to obtain MRIB information.|
|FEA→FIB2MRIB||to propagate the FIB updates to the FIB2MRIB module.|
|FIB2MRIB→RIB||to add/delete FIB-derived MRIB entries.|
|StaticRoutes→RIB||to add/delete MRIB-specific static routes.|
In the next subsections we describe those interactions.
The MFEA module needs to interact with the FEA module to obtain the vif-related information, and to keep track if there is any change. If there is any change in the vif information, the MFEA saves that information locally and propagates it to the modules that have registered interest with the MFEA (e.g., MLD/IGMP and PIM-SM).
PIM-SM needs to interact with the MFEA module for the following:
On startup, the PIM module registers with the MFEA. As part of this registration, the MFEA sends information about existing interfaces on the system, and the current unicast forwarding information. If any of this information changes later, the MFEA sends the appropriate update to PIM. In addition, the MFEA creates the appropriate state to send and receive PIM control packets. After the PIM module receives the network interface information, it instructs the MFEA to start PIM operation on selected interfaces. After that the PIM module can send and receive PIM control packets on those interfaces. Also, it can send requests to the MFEA to add, delete or modify multicast forwarding entries in the kernel. Note that the default solution in case of UNIX PIM kernel uses one more signal from the kernel to the user-level daemon: WHOLEPKT messages. Those messages are multicast packets that are suppose to be encapsulated in a PIM-SM Register message and sent to the RP. The encapsulation mechanism must know the RP address (in general, per multicast group address), threfore to avoid putting the RP addresses in the kernel (which would also change the traditional UNIX multicast API), the default solution is to use user-level encapsulation. To improve performance, the PIM encapsulation should be done in the kernel, and then we need a signaling mechanism between the kernel and the user process managing the kernel multicast forwarding entries. That mechanism would deal with installing the appropriate information needed by kernel-level Register encapsulation. Currently (July 2008), only recent releases of DragonFlyBSD, FreeBSD, NetBSD, and OpenBSD do support such kernel-level encapsulation.
PIM-SM needs to interact with the MLD/IGMP module to receive multicast membership information about local members. On startup, the PIM module registers with the MLD/IGMP module by expressing interest in tracking multicast membership on selected network interfaces. After that, if the multicast membership on any of the selected interfaces changes, the MLD/IGMP module informs the PIM module about the change: e.g., “add membership for group 18.104.22.168 on interface vif0”. Note that control packets used for multicast debugging such as MRINFO or multicast traceroute use IGMP as the network protocol. Hence, if such messages are received by the MLD/IGMP module, the appropriate information should be sent to the PIM module. Currently (July 2008), the handling of such messages is not implemented yet.
The PIM module needs to interact with the RIB module to obtain the MRIB information, and to keep track if there is any change.
Typically, the MRIB information inside the RIB module will be populated by entries from the unicast routing protocols running as part of a XORP router. For example, if Multiprotocol BGP is running on the system, it can be used to supply routes that are multicast-only; those routes then will be stored in the MRIB. However, if a XORP router is used only for multicast routing, and if there are no XORP unicast routing protocols to supply the MRIB routes, then by default we would use the existing unicast forwarding entries from the FIB in the underlying system. Those entries may have been statically installed on system startup, or may be managed by a non-XORP unicast routing implementation. The FIB2MRIB module is used to receive the FIB updates from the FEA. It also receives network interface related information from the FEA. The FIB information is updated by the network interface related information, (e.g., FIB entries that use disabled interfaces are withdrawn), and the result is propagated as MRIB entries to the RIB module.
The StaticRoutes module is used to store and manipulate the static unicast and multicast routes. A configured static route is added directly to the StaticRoutes module. This information is updated by the network interface related information obtained from the FEA (not shown on Multicast Modules Interaction figure above). For example, static routes that use disabled interfaces are withdrawn. The result is propagated as URIB or MRIB entries to the RIB module.