Lecture 8: Intradomain routing

Up until now, we have talked about the top two layers of the networking stack: the application and transport layers. This week we’ll start with the routing layer. This is the part of the network that gets packets from one end of the network to the other through a sequence of routers, called a path. The job of the routing layer is to find such a path between and source host and any destination host (routing), and then to send packets along this path (forwarding). These two functions: routing and forwarding, correspond to the two planes that are present on every router. The control plane handles routing and finds paths between any two hosts/routers in the network. The data plane actually sends (or forwards) packets along these pre-computed paths when a new packet shows up at the router. The interface between these two planes is the forwarding (or routing) table, which is essentially a look-up table (or a dictionary or hashmap or map or hash table depending on which term you prefer). Let’s look at the content of the look-up tables at any particular router. The keys in this look-up table are the destination addresses of the hosts in the network. The values in this look-up table are the output ports of this router on which a packet must be sent out so that the packet eventually finds its way to the destination. These output ports correspond to the next hop along the path from the source to the destination. Figure 1 shows the forwarding tables for four nodes connected in a straight line topology. The control plane writes entries (destination addresses and next hops) into the forwarding tables, while the data plane reads these entries by looking up the next hop for the destination address carried by the incoming packet. These two planes operate at very different timescales. The control plane needs to update or rewrite the forwarding tables roughly every time the topology changes because a new router joined the topology or an old one failed. The data plane, on the other hand, needs to read the forwarding tables on every incoming packet at the router. The rate of topology changes is typically much lower than the packet rate at a router. For instance, for a router that supports 1 Tbit/s of aggregate forwarding capacity, this translates into 1B packets/s, assuming 1000 bit packets. That’s a packet every 1 ns in the data plane. Topology changes, on the other hand, are relatively infrequent. Even in large networks (e.g., the ones inside a large company like Google, Microsoft, or Facebook), topology changes happen at the rate of at most a 1000 times a second. That’s a change every 1 ms in the control plane. Hence the timescales differ by six orders of magnitude. As a result, the data and control planes are implemented very differently. Because the data plane operates at such small time scales, it is built using dedicated hardware that is specialized for table lookups. This hardware is called a switching chip or a switching ASIC (Application Specific Integrated Circuit). The control plane is implemented using a general-purpose CPU like an Intel processor because that is sufficient for the rates required by the control plane. Keeping with this control and data plane dichotomy, our discussion of the routing layer will also be divided into two parts. This week we’ll look at the control plane, and the week after the midterm, we’ll look at the data plane. Let’s begin with the control plane. The job of the control plane is to compute the forwarding tables shown in Figure 1 and write them into the router. The job of the control plane can be summarized by a routing algorithm, which computes paths between any two hosts/routers within a network. We’ll now discuss two specific routing algorithms.