Lecture 11: Packet forwarding

This week we’ll talk about the data plane. Recall that the routing layer broadly consists of two parts: (1) the control plane that computes routes using distributed protocols such as the distance vector and link state protocols, and (2) the data plane that forwards packets based on routes computed by these protocols. The control plane kicks in every time the topology changes (e.g., a new router boots up, a router goes down, or a link cost is modified), while the data plane operates on every packet to look up the packet’s destination address and determine its output port. The interface between the control and data planes is the forwarding table, which is a map/dictionary/hashtable that maps destination IP addresses to the output port connecting to the next hop on the way to that destination. The control plane writes into this forwarding table, while the data plane reads from it. Because the data plane operates on a per-packet basis, it typically needs to be much faster than the control plane. On a high-speed 1 Tbit/s router today, assuming a 1000 bit packet, a packet arrives every 1 ns. It is quite hard to build a lookup table supporting 1 B lookups per second using an off-the-shelf CPU today. Hence, the data plane is built out of dedicated chips (called switching chips), which are essentially extremely large digital circuits within a single package that have been specialized to packet forwarding. The topic of this lecture will be packet forwarding: how to perform lookups at these high speeds when a new packet arrives every nanosecond. Then, in the next lecture, we’ll deal with packet buffering and scheduling, which is the task of storing packets and selecting packets for transmission when more than one input port simultaneously wants to send a packet to the same output port. OK. Now for packet forwarding. You’re given an IP address corresponding to the destination this packet must get to. The routing protocol has already computed a forwarding table that contains this destination. How do you lookup this IP address in the forwarding table to determine the port number of the next hop on the way to this IP address? And how do you do it quickly? We’ll consider two variants of this problem. The first is exact matching: a destination IP address is matched against every IP address in the forwarding table to find a unique next hop port—assuming there are no duplicates in the forwarding table. While this works, and is in fact how switches at the MAC layer operate, it suffers from a serious problem. It requires each router to store the next hop for every public IP address in its routing table. With 32-bit IP addressess, this requires every router to store around 4 billion forwarding table entries, which would take up several GBytes of router memory. Memories that are cheap enough to store several GBytes (e.g., DRAM memory found in your processor memories) are also typically too slow to support 1 billion lookups per second. Besides, this requires every router to know of every other router/end-host as part of its routing protocol, which is a serious problem. We have looked at the solution to this in the lecture on inter-domain routing: hierarchy. Hierarchy allows us to separate inter and intradomain routing by allowing each inter-domain router to only know about other inter-domain routers and each intra-domain router to only know about other routers within its own domain. This separation reduces (1) the volume of traffic generated by the routing protocol because each routing protocol (whether inter or intradomain) involves fewer entities, and (2) the size of the forwarding tables, again because each router needs to forward packets to fewer other routers. But, as a result of this separation, a router deep inside the network only knows how to get to the network of a particular destination address. It does not know how to get to the address itself. These networks are represented using a format similar to IP addresses, but with the allowance that some bits can be wildcards. For instance, 8.8.8.* represents the network with IP addresses between 8.8.8.1 and 8.8.8.255, because the * can