Ethernet Performance

Ethernet Performance

Now let us briefly examine the performance of classic Ethernet under conditions of heavy and constant load, that is, with k stations always ready to transmit. A rigorous analysis of the binary exponential backoff algorithm is complicated. Instead, we will follow Metcalfe and Boggs (1976) and assume a constant retransmission probability in each slot. If each station transmits during a contention slot with probability p, the probability A that some station acquires the channel in that slot is 

                                                                 A = kp(1 − p) ^{k − 1}                                                                (4-5)

A is maximized when p = 1/k, with A → 1/e as k → ∞. The probability that the contention interval has exactly j slots in it is A(1 − A) ^{j − 1}, so the mean number of slots per contention is given by

  Since each slot has a duration 2τ, the mean contention interval, w, is 2τ/A. Assuming optimal p, the mean number of contention slots is never more than e, so w is at most 2τe ∼∼ 5.4τ.

If the mean frame takes P sec to transmit, when many stations have frames to send,

Here we see where the maximum cable distance between any two stations enters into the performance figures. The longer the cable, the longer the contention interval, which is why the Ethernet standard specifies a maximum cable length.It is instructive to formulate Eq. (4-6) in terms of the frame length, F, the network bandwidth, B, the cable length, L, and the speed of signal propagation, c, for the optimal case of e contention slots per frame. With P = F/B, Eq. (4-6) becomes

When the second term in the denominator is large, network efficiency will be low. More specifically, increasing network bandwidth or distance (the BL product) reduces efficiency for a given frame size. Unfortunately, much research on network hardware is aimed precisely at increasing this product. People want high bandwidth over long distances (fiber optic MANs, for example), yet classic Ethernet implemented in this manner is not the best system for these applications. We will see other ways of implementing Ethernet in the next section.

In Fig. 4-16, the channel efficiency is plotted versus the number of ready stations for 2τ = 51.2 μsec and a data rate of 10 Mbps, using Eq. (4-7). With a 64- byte slot time, it is not surprising that 64-byte frames are not efficient. On the other hand, with 1024-byte frames and an asymptotic value of e 64-byte slots per contention interval, the contention period is 174 bytes long and the efficiency is 85%. This result is much better than the 37% efficiency of slotted ALOHA.

It is probably worth mentioning that there has been a large amount of theoretical performance analysis of Ethernet (and other networks). Most of the results should be taken with a grain (or better yet, a metric ton) of salt, for two reasons.

First, virtually all of the theoretical work assumes Poisson traffic. As researchers have begun looking at real data, it now appears that network traffic is rarely Poisson. Instead, it is self-similar or bursty over a range of time scales (Paxson and Floyd, 1995; and Leland et al., 1994). What this means is that averaging over long periods of time does not smooth out the traffic. As well as using questionable models, many of the analyses focus on the ‘‘interesting’’ performance cases of abnormally high load. Boggs et al. (1988) showed by experimentation that Ethernet works well in reality, even at moderately high load.

Switched Ethernet

Ethernet soon began to evolve away from the single long cable architecture of classic Ethernet. The problems associated with finding breaks or loose connections drove it toward a different kind of wiring pattern, in which each station has a dedicated cable running to a central hub. A hub simply connects all the attached wires electrically, as if they were soldered together. This configuration is shown in Fig. 4-17(a).

The wires were telephone company twisted pairs, since most office buildings were already wired this way and normally plenty of spares were available. This reuse was a win, but it did reduce the maximum cable run from the hub to 100 meters (200 meters if high quality Category 5 twisted pairs were used). Adding or removing a station is simpler in this configuration, and cable breaks can be detected easily. With the advantages of being able to use existing wiring and ease of maintenance, twisted-pair hubs quickly became the dominant form of Ethernet.

However, hubs do not increase capacity because they are logically equivalent to the single long cable of classic Ethernet. As more and more stations are added, each station gets a decreasing share of the fixed capacity. Eventually, the LAN will saturate. One way out is to go to a higher speed, say, from 10 Mbps to 100 Mbps, 1 Gbps, or even higher speeds. But with the growth of multimedia and powerful servers, even a 1-Gbps Ethernet can become saturated.

Fortunately, there is an another way to deal with increased load: switched Ethernet. The heart of this system is a switch containing a high-speed backplane that connects all of the ports, as shown in Fig. 4-17(b). From the outside, a switch looks just like a hub. They are both boxes, typically with 4 to 48 ports, each with a standard RJ-45 connector for a twisted-pair cable. Each cable connects the switch or hub to a single computer, as shown in Fig. 4-18. A switch has the same advantages as a hub, too. It is easy to add or remove a new station by plugging or unplugging a wire, and it is easy to find most faults since a flaky cable or port will usually affect just one station. There is still a shared component that can fail—the switch itself—but if all stations lose connectivity the IT folks know what to do to fix the problem: replace the whole switch.

 

Inside the switch, however, something very different is happening. Switches only output frames to the ports for which those frames are destined. When a switch port receives an Ethernet frame from a station, the switch checks the Ethernet addresses to see which port the frame is destined for. This step requires the switch to be able to work out which ports correspond to which addresses, a process that we will describe in Sec. 4.8 when we get to the general case of switches connected to other switches. For now, just assume that the switch knows the frame’s destination port. The switch then forwards the frame over its high-speed backplane to the destination port. The backplane typically runs at many Gbps, using a proprietary protocol that does not need to be standardized because it is entirely hidden inside the switch. The destination port then transmits the frame on the wire so that it reaches the intended station. None of the other ports even knows the frame exists.

What happens if more than one of the stations or ports wants to send a frame at the same time? Again, switches differ from hubs. In a hub, all stations are in the same collision domain. They must use the CSMA/CD algorithm to schedule their transmissions. In a switch, each port is its own independent collision domain. In the common case that the cable is full duplex, both the station and the port can send a frame on the cable at the same time, without worrying about other ports and stations. Collisions are now impossible and CSMA/CD is not needed. However, if the cable is half duplex, the station and the port must contend for transmission with CSMA/CD in the usual way.

A switch improves performance over a hub in two ways. First, since there are no collisions, the capacity is used more efficiently. Second, and more importantly, with a switch multiple frames can be sent simultaneously (by different stations). These frames will reach the switch ports and travel over the switch’s backplane to be output on the proper ports. However, since two frames might be sent to the same output port at the same time, the switch must have buffering so that it can temporarily queue an input frame until it can be transmitted to the output port. Overall, these improvements give a large performance win that is not possible with a hub. The total system throughput can often be increased by an order of magnitude, depending on the number of ports and traffic patterns.

The change in the ports on which frames are output also has security benefits. Most LAN interfaces have a promiscuous mode, in which all frames are given to each computer, not just those addressed to it. With a hub, every computer that is attached can see the traffic sent between all of the other computers. Spies and busybodies love this feature. With a switch, traffic is forwarded only to the ports where it is destined. This restriction provides better isolation so that traffic will not easily escape and fall into the wrong hands. However, it is better to encrypt traffic if security is really needed.

Because the switch just expects standard Ethernet frames on each input port, it is possible to use some of the ports as concentrators. In Fig. 4-18, the port in the upper-right corner is connected not to a single station, but to a 12-port hub instead. As frames arrive at the hub, they contend for the ether in the usual way, including collisions and binary backoff. Successful frames make it through the hub to the switch and are treated there like any other incoming frames. The switch does not know they had to fight their way in. Once in the switch, they are sent to the correct output line over the high-speed backplane. It is also possible that the correct destination was one on the lines attached to the hub, in which case the frame has already been delivered so the switch just drops it. Hubs are simpler and cheaper than switches, but due to falling switch prices they have become an endangered species. Modern networks largely use switched Ethernet. Nevertheless, legacy hubs still exist.

Frequently Asked Questions