Next Step for Ethernet Switch Design
Ethernet switches, like most computing devices, are getting smarter. A few years ago, they were dumb, “plug-n-play” sort of things—they had a small number of interfaces and they performed a simple role. They would receive a packet of data, look at the Ethernet address, and send the packet out on one or more of the other interfaces. They could learn the source address from a received packet, and use that knowledge to decide where to send a future packet. But if they didn’t know the address, or it was a “special” address (broadcast or multicast) they had to send it out on every interface. This operation is nice and simple, and works—but isn’t very efficient.
Over the past 10 or 15 years, things have moved on. Firstly, the speed of transmission has increased, from 10Mbps, to 100Mbps to Gigabit, then 10GbE, and we’re now talking 40GbE or even 100GbE. But as well as being faster, they are also smarter. Switches, generally, have moved from being “unmanaged” to “managed.” This management allows them to do a number of significant things—mainly aimed at making them handle traffic in a more efficient way. Examples of this are:
- Knowledge of how multicast works
- Ability to create Virtual LANs (VLANS) to provide logical separation on one physical network
- Control and monitoring—things like getting statistics from each port
- Understanding Layer 3 addresses—doing things that “routers” used to
- Supporting protocols for protecting the network from failures
In order to undertake this management, the hardware components in the switch have changed. At the heart of an unmanaged switch is, basically, a complex integrated circuit, called a “switch fabric.” There are also components related to the physical transmission and reception of data, but the “smarts” are all in the switch fabric (which is actually an ASIC, if you like to get into the details). This switch fabric is great at taking simple decisions very fast on where to send a packet, but can’t make complex decisions.
Unmanaged vs. managed switches
When we move from unmanaged to managed switches, we add another important component to this mix—a general processor (CPU) of some sort. Typically, these are low-powered, general purpose CPUs, maybe PowerPC based. Their job is to understand the things the switch fabric doesn’t—handling the configuration, communicating with the user and so on. So, the fabric does the simple, fast stuff; the CPU does the complex, slow stuff.
This picture has remained quite similar for more than a decade, but recent advances from the producers of the fabrics have changed it subtly yet significantly. Using system-on-a-chip (SoC) techniques, we can now get switch fabrics that incorporate a CPU core into the one integrated circuit. Typically, these CPU cores are something like ARM or MIPS CPUs, and have a direct interface to the fabric.
What’s the big advantage with these? Well, there are a few, but the two major ones are:
- Lower power
- Smaller board space
At GE, we’ve already developed a new switch using such an integrated fabric / CPU, and it’s at the working prototype stage. We’re porting our OpenWare switch management software to run in this environment, and it will be shipping very soon. The first customer for this has a requirement for a switch to add features into a legacy form factor, and this is the perfect first step for this technology. The customer needed a managed switch to undertake more complex tasks, but it had to fit in the same board size as an unmanaged switch.
The next step for this is a new series of rugged Ethernet switches, currently being referred to as the RES-300 family. Our site in Towcester is working on the design for these. These will be our introduction of this integrated fabric/CPU into the military environment. Now, we will be able to enable our customers to handle more traffic in more complex setups, with less size, weight, power and cost—the famous “reduced SWaP-C.”
Watch this space!