Chapter |Capacity|: Allocating Capacity
We know from Chapter |Intro| that statistical multiplexing is the foundation for sharing network resources in a best-effort network, but it isn't the full story about how network resources are allocated. This chapter fills in the remaining details, looking at how networks deal with varying traffic loads at both per-node and network-wide levels.
Each node has two resources that need attention: link bandwidth and buffer capacity. (The third resource described in Section |Intro|.4.3—internal forwarding capacity—could in principle be managed dynamically, but in practice is not.) With respect to link bandwidth, statistical multiplexing is the starting point, with packets scheduled for transmission based on demand. But there are ways to augment statistical multiplexing both to be more fair (i.e., to keep greedy flows from starving well-behaved flows) and to ensure that important packets don't get stuck behind unimportant packets.[1] Such algorithms, known as packet schedulers, effectively allocate link capacity by deciding which packet to send next.
| [1] | The idea that some packets are more important than others raises a set of issues to be discussed later, such as who decides which packets are important. |
Packet scheduling is an aspect of buffer management—the scheduler decides which packet to dequeue—but there are additional techniques for managing packet queues. They are collectively known as Active Queue Management (AQM), and they generally decide which packet(s) to drop as the buffer fills up. AQM plays an especially important role in the Internet because dropping a packet is an indirect way for a network node to "signal" an edge host—the entity responsible for injecting the packet into the network—that something is amiss.
Ultimately, responsibility for managing congestion is shared between routers inside the network (they detect the onset of congestion) and edge hosts (they adjust their sending rates in response to congestion). This chapter focuses on the router side of this interaction, and we cover the edge host side of this interaction in Chapter |CC|. For the purpose of this chapter, all we need to know about edge hosts is that they react to whatever signal the routers send them by adapting the rate at which they inject packets into the network.
These node-level mechanisms have to make local decisions in real time, but there are also network-wide actions that dictate how to place that traffic onto links and nodes in the first place. This network-wide decision making process is known as Traffic Engineering (TE), and it comes into play when per-node capacity is chronically under-provisioned due to long-term changes in the volume of traffic flowing between various nodes. TE was originally an off-line process, requiring administrator intervention, but it is increasingly being done as an automated (but coarse-grained) part of a network's overall approach to capacity management.