Basic Networking Concepts
What is a data network? At its most abstract, a data network is a set of nodes connected by links. In the context of data networks, the nodes are routers, LAN switches, WAN switches, add-drop multiplexers (ADMs), and the like, connected by links from 64 Kb DS0 circuits to OC192 and 10 gigabit Ethernet.
One fundamental property of data networks is multiplexing. Multiplexing allows multiple connections across a network to share the same transmission facilities. Two main types of multiplexing to be concerned with are
Other kinds of multiplexing, such as frequency-division multiplexing (FDM) and wavelength-division multiplexing (WDM) are not discussed here.
TDM
Time-division multiplexing is the practice of allocating a certain amount of time on a given physical circuit to a number of connections. Because a physical circuit usually has a constant bit rate, allocating a fixed amount of time on that circuit translates directly into a bandwidth allocation.
A good example of TDM is the Synchronous Optical Network (SONET) hierarchy. An OC192 can carry four OC-48s, 16 OC-12s, 64 OC-3s, 192 DS-3s, 5376 DS-1s, 129,024 DS-0s, or various combinations. The Synchronous Digital Hierarchy (SDH) is similar.
TDM is a synchronous technology. Data entering the network is transmitted according to a master clock source so that there's never a logjam of data waiting to be transmitted.
The fundamental property of TDM networks is that they allocate a fixed amount of bandwidth for a given connection at all times. This means that if you buy a T1 from one office to another, you're guaranteed 1.544 Mbps of bandwidth at all times—no more, no less.
TDM is good, but only to a point. One of the main problems with TDM is that bandwidth allocated to a particular connection is allocated for that connection whether it is being used or not. Thirty days of T1 bandwidth is roughly 4 terabits. If you transfer less than 4 terabits over that link in 30 days, you're paying for capacity that you're not using. This makes TDM rather expensive. The trade-off is that when you want to use the T1, the bandwidth is guaranteed to be available; that's what you're paying for.
Statistical Multiplexing
The expense of TDM is one reason statistical multiplexing technologies became popular. Statistical multiplexing is the practice of sharing transmission bandwidth between all users of a network, with no dedicated bandwidth reserved for any connections.
Statistical multiplexing has one major advantage over TDM—it's much cheaper. With a statmux network, you can sell more capacity than your network actually has, on the theory that not all users of your network will want to transmit at their maximum bit rate at the same time.
There are several statmux technologies, but the three major ones in the last ten years or so have been
MPLS is a fourth type of statmux technology. How it fits into the picture is explained later in this chapter.
Statmux technologies work by dividing network traffic into discrete units and dealing with each of these units separately. In IP, these units are called packets; in Frame Relay, they're called frames; in ATM, they're called cells. It's the same concept in each case.
Statmux networks allow carriers to oversubscribe their network, thereby making more money. They also allow customers to purchase network services that are less expensive than TDM circuits, thereby saving money. A Frame Relay T1, for example, costs far less than a TDM T1 does. The ratio of bandwidth sold to actual bandwidth is the oversubscription ratio. If you have an OC-12 backbone and you sell 24 OC-3s off of it, this is a 6:1 oversubscription ratio. Sometimes, this number is expressed as a percentage—in this case, 600 percent oversubscription.
Issues That Statmux Introduces
Statmux introduces a few issues that don't exist in TDM networks. As soon as packets enter the network asynchronously, you have the potential for resource contention. If two packets enter a router at the exact same time (from two different incoming interfaces) and are destined for the same outgoing interface, that's resource contention. One of the packets has to wait for the other packet to be transmitted. The packet that's not transmitted needs to wait until the first packet has been sent out the link in question. However, the delay encountered because of simultaneous resource contention on a non-oversubscribed link generally isn't that big. If 28 T1s are sending IP traffic at line rate into a router with a T3 uplink, the last IP packet to be transmitted has to wait for 27 other IP packets to be sent.
Oversubscription greatly increases the chance of resource contention at any point in time. If five OC-3s are coming into a router and one OC-12 is going out, there is a chance of buffering because of oversubscription. If you have a sustained incoming traffic rate higher than your outgoing traffic capacity, your buffers will eventually fill up, at which point you start dropping traffic.
There's also the issue of what to do with packets that are in your buffers. Some types of traffic (such as bulk data transfer) deal well with being buffered; other traffic (voice, video) doesn't. So you need different packet treatment mechanisms to deal with the demands of different applications on your network.
Statmux technologies have to deal with three issues that TDM doesn't:
Buffering Queuing Dropping
Dealing with these issues can get complex.
Frame Relay has the simplest methods of dealing with these issues—its concepts of committed information rate (CIR), forward and backward explicit congestion notification (FECN and BECN), and the discard eligible (DE) bit.
IP has DiffServ Code Point (DSCP) bits, which evolved from IP Precedence bits. IP also has random early discard (RED), which takes advantage of the facts that TCP is good at handling drops and that TCP is the predominant transport-layer protocol for IP. Finally, IP has explicit congestion notification (ECN) bits, which are relatively new and as of yet have seen limited use.
ATM deals with resource contention by dividing data into small, fixed-size pieces called cells. ATM also has five different service classes:
CBR (constant bit rate) rt-VBR (real-time variable bit rate) nrt-VBR (non-real-time variable bit rate) ABR (available bit rate) UBR (unspecified bit rate)
Statmux Over Statmux
IP was one of the first statmux protocols. RFC 791 defined IP in 1981. The precursor to IP had been around for a number of years. Frame Relay wasn't commercially available until the early 1990s, and ATM became available in the mid-1990s.
One of the problems that network administrators ran into as they replaced TDM circuits with Frame Relay and ATM circuits was that running IP over FR or ATM meant that they were running one statmux protocol on top of another. This is generally suboptimal; the mechanisms available at one statmux layer for dealing with resource contention often don't translate well into another. IP's 3 Precedence bits or 6 DSCP bits give IP eight or 64 classes of service. Frame Relay has only a single bit (the DE bit) to differentiate between more- and less-important data. ATM has several different service classes, but they don't easily translate directly into IP classes. As networks moved away from running multiple Layer 3 protocols (DECnet, IPX, SNA, Apollo, AppleTalk, VINES, IP) to just IP, the fact that the Layer 2 and Layer 3 contention mechanisms don't map well became more and more important.
It then becomes desirable to have one of two things. Either you avoid congestion in your Layer 2 statmux network, or you find a way to map your Layer 3 contention control mechanisms to your Layer 2 contention control mechanisms. Because it's both impossible and financially unattractive to avoid contention in your Layer 2 statmux network, you need to be able to map Layer 3 contention control mechanisms to those in Layer 2. This is one of the reasons MPLS is playing an increasingly important part in today's networks—but you'll read more about that later.
|
