Lecture 12: Input and output queueing

We’ll continue our discussion of the data plane this lecture, but focus on the queueing subsystem within a router’s data plane. But first, what is the source of queues within a router’s data plane? Queues arise because multiple input ports on a router might want to send packets to the same output port. A number of scenarios can result in this type of multiple-input-single-output (also known as fanin/incast) behavior. One example is that of several client machines talking to a single server machine as in the case of a flash crowd to a very popular server. Another example scenario is what happens internally at a search provider like Google when a user issues a search query: the search query is received by a frontend Google server, which then delegates the query to several worker machines, where each worker machine runs the user’s search query on a subset of the Internet’s web pages that Google has previously crawled. These worker machines all return their query results to the frontend server at roughly the same time (because they are all given the same time budget to work on their queries). This causes a fanin or incast at the frontend server. When multiple inputs send packets to the same output at roughly the same time, we need to deal with the fact that the input arrival rate is higher than the output departure rate—at least temporarily. The standard solution is to transmit as many packets as the output port can transmit and queue up the rest somewhere for later transmission. If the situation persists, the queue might overflow and drop packets. This is what causes TCP’s AIMD algorithm to cut its window in half. In this lecture, we’ll look at a few architectures for organizing these queues within a router and their associated tradeoffs. We’ll finally discuss how technology trends determine the right choice of the architecture for the queueing subsystem.