Ethernet is everywhere. That’s no surprise. In many ways, it’s as close to perfect as a technology can be in terms of its scalability, its flexibility, its simplicity and its low cost. From 10 Mbits/second through 100 Mbits/second and on to Gbit Ethernet, 10 Gigabit Ethernet–and beyond?–demonstrates its scalability (and each successive generation has been backward-compatible with its predecessor). It has enabled networks in a host of sizes and configurations–and is, increasingly, finding application as a backplane interconnect. It’s a technology with which virtually every computer user feels comfortable: what, after all, could be simpler than plugging one end of a cable into the back of a PC, and the other into a router? And the cables, terminators, switches, multiplexes, modems and routers that give shape to an Ethernet network can be had for the equivalent of mere pennies.
Ethernet has definitely made its mark in the military realm. Rugged Ethernet Switch products are available in many form factors, including VME, VXS and VPX. Switched Ethernet technology is finding its way into numerous programs, both new and upgrades. Switched Ethernet is being used, for example, as an interconnect for the upgraded electronics on BAE Systems’ Bradley Fighting Vehicle Program (Figure 1). The Ethernet Switch Unit (ESU) in the A3 Bradley Combat Systems vehicle functions as a router and a switch, making local forwarding decisions to devices operated in the vehicle’s LAN. The A3 upgrade version of the Bradley features an advanced digital architecture that integrates communications equipment, digital sensors, battle management systems, embedded diagnostic and training systems.
What is perhaps surprising, however, is the way in which Ethernet has succeeded despite the accusations leveled against it. Some of the criticisms are perfectly valid–but others are much less so. Take, for example, one of the key concepts underlying Ethernet in action: CSMA/CD (Carrier Sense Multiple Access with Collision Detection). The principle of CSMA/CD was that there was a shared medium (originally a big, thick yellow cable) that everyone on the network would try to use. If two nodes tried to use it at the same time, they ‘detected’ the ‘collision’, backed off, and tried again.
Non-Determinism Problem
The time delay before retrying was controlled by random and exponential elements. CSMA/CD was a powerful concept, but it led to one of the most basic criticisms: that transmission times were unpredictable and not deterministic, meaning that network design became as much a matter of art as of science. There was no 100% guarantee that any transmission would happen within x milliseconds–or even, in fact, within x seconds. Today’s reality, however, is that this ‘non-determinism’ caused by access to shared media is no longer true in a world in which full duplex, switched Ethernet networks are the norm. However, the role of CSMA/CD in Ethernet’s alleged non-determinism remains one of the most common criticisms.
It is true that there are other causes of the non-determinism that occurs even in full-duplex, switched Ethernet networks–but the practical issues that result from this are very few, and pragmatic answers, such as priority queuing, are solving those. This is a good example of how Ethernet technology has widened and adapted to suit a more diverse set of transmission needs over the years. In its early years, Ethernet traffic would typically have seen a dumb terminal providing user input/output–that may have been field-oriented in order to improve efficiency–carried over a sophisticated network architecture such as DECnet. Today, it is more likely to be Google search results, encapsulated voice, or even captured video streams.
One area where Ethernet has been criticized without good reason is when designers fail to distinguish between the Ethernet technology itself, and the protocols that are commonly used above it. Much has been said about the CPU processing overhead required to calculate IP checksums, or the complexity of packet-window sequencing in TCP. However, these have nothing to do with Ethernet itself.
Protocol Layering
While IP is the most commonly used protocol over an Ethernet network, the arrangement is not an exclusive one, thanks to the simple–elegant, even–concept of protocol layering. Ethernet provides a simple frame-level transmission/reception mechanism–the bottom two layers of the once-famous OSI Reference Model. It can be used (as most installations do) to carry the IP protocol (Layer 3). Then, above IP, TCP might be run, or UDP, Voice-over-IP (VOIP), or any of a number of alternatives (Layers 4 through 7).
That said, it is equally possible to use Ethernet framing to carry any other protocol that can be split into frames (and most can). So, encapsulating ATM frames straight into Ethernet frames is a common solution for certain telecoms situations. Of course, in this case, all the added functionality that a higher-layer protocol like IP brings is lost–but this may be appropriate depending on the network scenario.
There are still those who maintain negative views about the overheads of the TCP and IP layers–but the issues associated with those overheads are now being addressed. There is now significant support in specialist hardware for the off-loading of functions like segmentation and checksum calculations–meaning that the TCP/IP protocol stack is no longer the CPU-killer it was once made out to be.
Recently, Ethernet signaling has found substantial favor–along with the likes of Serial RapidIO and PCI Express–in backplanes such as those that can be found in CompactPCI 2.16, ATCA, MicroTCA, VSP, VPX and so on. This can be seen as a further acknowledgement of the simplicity, reliability and cost benefits of Ethernet. In this circumstance, the designer is not looking for the campus-size reach that Ethernet was originally designed for, but a tried-and-trusted transmission mechanism. Given that Ethernet is forecast to soon reach transmission speeds of 40 Gbits/s or even 100 Gbits/second, it is clear that Ethernet is capable of providing backplane speeds that current processors would be hard pressed to fill with meaningful data.
There will, of course, always be comparisons made between Ethernet and other backplane technologies such as Serial Rapid IO, PCI Express and so on. The decision about which of these is most suitable is complex, and is normally driven by the particular application: Ethernet is, however, proving itself suitable for a growing number of applications.
It is also important to bear in mind that the various serial switched fabrics currently achieving a high profile are not mutually exclusive. Each has its strengths and weaknesses, and the best systems designs leverage those strengths. It’s possible, for instance, to envisage a high-performance system in which sensors are connected to high-speed ADCs, which in turn are connected via PCI Express, with Serial RapidIO for node-to-node transfers within a back-end multiprocessing system. The results could then be relayed via Ethernet to a host computer for visualization (Figure 2). That kind of highly typical application architecture is probably what Freescale had in mind when integrating all three fabrics into the 8640 and 8640D processors.
Internet-Style Protocols
It is important to acknowledge that the majority of casual references to Ethernet also bring with them the inference of the rest of the ‘Internet-style’ protocols such as IP, TCP, UDP, FTP and HTTP. Many of the backplane users of Ethernet are also using these higher layer protocols, and the reason for choosing Ethernet as a backplane transmission method is often its suitability for carrying this type of traffic. The benefit, of course, is that traffic across the backplane can be identical to traffic between backplanes–or even around the world.
This ‘geographic transparency’ to the software applications can be attractive. The system designer doesn’t need to care if the environment is a test situation, where video capture data is from a camera balanced on top of a coffee-cup on a desk, or a deployment situation, where it’s in an unmanned drone over some arid desert.
Of course, a fundamental device in modern Ethernet is the switch. Switches did not exist in the original Ethernet concept. First, there was the idea of a ‘repeater’ to amplify the signals on those lengths of thick, yellow cable. There was also the concept of the ‘bridge’ that enabled the connection of two segments of cable, while limiting traffic between the two. The advent of ‘Cheapernet’ (10Base5) saw the introduction of the hub, allowing the connection of multiple segments. The Ethernet switch is, in effect, a combination of the concept of the hub and the concept of the bridge.
Ethernet switches come in a range of capabilities, from the ‘plug-n-play’ (sometime called ‘dumb’) Layer 2 through ‘managed’ Layer 2–such as GE Fanuc’s RM983RC 6U VME 12- or 24-port switch (Figure 3) to ‘fully managed’ Layer 3. The system designer needs to ensure that the right type of switch is chosen, bearing in mind the needs of the application. A Layer 2 switch makes all its decisions on the basis of information contained in the Layer 2 header–the so-called MAC (Media Access Control) layer, the highest layer that Ethernet knows anything about. Layer 3 switches are able to make more sophisticated decisions based on information from the IP layer–and, in some cases, on information from above the IP layer such as TCP.
Ethernet Switches as Uplinks
Switches that reside on Ethernet backplanes also often offer ‘uplinks’, allowing connection to other chassis, or to the outside world. These uplinks can be configured to simply look like the backplane ports, or to be in some way ‘special’. Some switches will allow configuration as Layer 3–the IP layer, making the switches take on some of the role of a router, and often controlling the traffic between the backplane and the outside world.
A general rule of thumb for switches might be that unmanaged Layer 2 switches are the simplest, and will almost always work, but may not give the level of efficiency required by the application. Managed Layer 2 switches can increase efficiency at the MAC level (for example, the ability to deal with VLANs and multicast handling). Managed Layer 3 switches provide the ultimate in configurability, and are often necessary for IP-based networks.
Ethernet has come a long way in its 30+ years of existence, and has demonstrated a remarkable ability to remain not only a relevant but also an essential element of today’s advanced networks. Its potential for high performance, its flexibility, its low cost and the huge ecosystem that now exists of hardware, software and expertise will ensure that it continues to be equally central to tomorrow’s networks.
GE Fanuc Intelligent Platforms
Charlottesville, VA.
(800) 368-2738.
[www.gefanucembedded.com].
