8.10 Multiple Line Card, Fully Distributed Architecture

In this IPS implementation scenario, each line card has its own control processor (see Figure 8.7). The control processors on the line cards communicate with each other through a control interface on the backplane such as PCI or Fast Ethernet. There is no single control card. Instead, the control function is distributed among the control processors on the line cards. For simplicity, the management interface is present on one of the line cards, so there is a master agent on the line card and subagents on other line cards. Management PDUs are routed to the line card with the master agent.

Each of the control processors runs the complete set of routing protocols. They communicate over the control interface with each other, so they effectively look like multiple routers in a box, with the backplane control interface acting like a LAN. So the control cards actually run the routing protocols over the backplane LAN. Since the routing functionality is now distributed, the control processors are not as heavily loaded as a single control card.

Using Figure 8.7 as an example, assume that the SNMP agent is located on Line Card 3, and the external manager communicates with the router on Line Card 1. Also, assume the variable value requested by the SNMP manager is obtained from Line Card 2. The SNMP packet is forwarded by Line Card 1 over the control backplane to Line Card 3. The SNMP agent on Line Card 3 communicates with the subagent on Line Card 2 to determine the value of the variable required by the external manager. This implies that the MIB views provided by the individual subagents are aggregated by the agent on Line Card 3, even though each of the Line Cards appears as a single router.

Figure 8.7: Fully distributed architecture.

While the distribution and MIB view abstraction is acceptable for tables, it is not as straightforward when we consider the routing protocols. If the OSPF protocol is to appear as a single entity to the external manager, the various instances on each of the line cards have to collaborate to provide the single OSPF instance view. This is conceptually similar to the split control plane view for OSPF 'Hello's, which involved two state machines cooperating through messages. The difference is that individual state machines are split across multiple boards but have to provide a single view for management.

Some systems avoid this problem altogether by providing the multiple OSPF instance view without having to combine the instances. However, the view is accessible only with the management agent and an instance ID-and not directly through the line card. This enables the administrator to configure the OSPF instances on a line card basis but at the same time not address them as individual routers. This is also called a virtual router approach. Some IP services boxes have used this architecture for scalability.

A variation on this is the multi-router approach, proposed recently. Here, an administrator can dynamically configure control and line card mappings, so that there are multiple physical routers in the same box. In this case, there is no need to have a unified management view for the routing protocols or cards, except to configure the mapping.