The evolution and future of Ethernet — is it the perfect predator?


Sharks are so deadly and efficient that they’ve been called “the perfect predator.” That’s why I find it fascinating that they’ve been the top predator in the oceans for literally hundreds of millions of years, and haven’t changed much since the extinction of the dinosaurs. Why not? They didn’t need to! Their basic design was so good that it only needed tweaking to survive while other species in water and on land died out (sorry, dinosaurs) or evolved so much you wouldn’t recognize them from their ancestors. (sorry, creationists)

Sound like any networking technologies you know? While I hesitate to admit this, I’ve been using Ethernet since before some of you were born. Ethernet may not be perfect, but it’s very much like the shark in that aside from improvement over time (higher bandwidth, new kinds of wires) it’s a lot like its ancestors. In fact, I have some “museum piece” computers at home running 10 MbitE that are talking to my 1 GigE hosts.

I won’t complain about how I used to walk to school every day (In the snow! Uphill! Both ways!), but I do remember implementing Ethernet over thick N-Series cable by connecting a DIX (Digital-Intel-Xerox standard) with “Vampire Clamps” and connecting the network card to the DIX via a transceiver cable. All that for 10 Mbits/second!

Today Ethernet runs over twisted pair and fiber, and it flows at up to 10 Gbits a second. Switches have taken the place of DIX and N-Series cables, and the interface technology is now extremely intelligent NIC cards. The protocol itself has been tweaked and standardized (thank you IEEE), but it’s alive and kicking. By comparison, when’s the last time you implemented any Token Ring? Have you upgraded your FDDI network to the latest version? Bought any XNS from Xerox lately?

Is 10GigE the end of the road?

10GigE brings us up to the bus speeds of PCI so I can’t go any faster in transferring data from the network card across the bus for processing in my computer. Does that mean 40GigE and 100GigE, both already being discussed, represent a case of a misplaced “bigger is better” mentality? No way, Jose!

What about new technologies like Infiniband, Fiber Channel, QsNet or Myrinet? Will one of them displace Ethernet? Only time will tell, but I personally think that in a couple years they will be footnotes in the Wikipedia page about 100GigE.

Here’s some of the reasons we’re always gonna need a bigger boat, and why I believe Ethernet will continue to be the networking technology of choice for many years to come.

  • The need for more at the core: The “tributary, stream, river, ocean” architecture allowed the Internet to become what it is today. Clearly applications and users who access the Internet don’t need the same high-speed network access that the core IP routers do. I surf the Internet with my iPhone as easily as I can with a super computer. The same hierarchical architecture is developing around messaging technology. So while it’s possible that the faster Ethernet standards may not be widely implemented in generic host computers, they will be immensely valuable at the core of the Internet, LANs and WANs, and will be critical for continuously unlocking new levels of performance in throughput-centric applications like (hardware-based) messaging.
  • Bandwidth is still the gating factor for some applications: With the deployment of 10 GigE, bandwidth isn’t the gating factor for most scenarios – it’s the computer. A great deal of effort is being spent offloading TCP/IP processing into network cards to reduce the processing overhead of the host computer, and there’s a lot of value in doing so. But at the end of the day computer memory is relatively slow, operating systems are stretched to the limit, and developers have taken to relying on faster hardware to keep their increasingly bloated software running. But there are applications where 10GigE is immensely valuable. Messaging, for example, has hit the wall when it comes to using generic hosts to run software-based brokers. If you eliminate the generic computer architecture and OS, your gating factor becomes the network. Solace’s new 10 GigE I/O blade implements all 7 layers of message processing in hardware in a single card, and interfaces with other specialized cards on the same box to minimize inter-card communications. Even then, the data path is implemented entirely in hardware within the box, so even when there are transfers between hardware elements they are not encumbered by an operating system.
  • This entire process mirrors the evolution of the Internet: The early Internet used UNIX hosts as routers, but eventually it took high performance hardware-based technology (routers and switches) to support its meteoric growth. We’re at the same stage for messaging—it takes hardware platforms to provide the services, and this hardware will eventually require the next generations of Ethernet to allow further expansion of messaging.

The Future of Ethernet

As the deployment of the hardware message router allows the volume of messages and hosts to increase, there will eventually be a need to aggregate the messages through higher volume and lower latency networks. Solace and other throughput-conscious vendors have built their products ready to accept and support 40GigE and even 100 GigE when they are available.
Building around a standard technology that can evolve and interact with earlier versions of itself represents the most logical and cost-effective path. That’s why Ethernet will continue to survive and evolve because, like the shark, it has a great basic design that can be improved with relatively minor modifications.