The IEEE wants to make idle or under-utilised Ethernet connections more energy efficient, which could mean huge electrical cost savings for large enterprises. The trick: finding a way to seamlessly throttle between 10Mbit/s and 10Gbit/s.

The standards outfit recently formed an Energy Efficient Ethernet (EEE) study group to explore how to do this. The idea is to save power in PCs and laptops (most of which ship with Gigabit cards now) when LAN links are idle, or not utilising full bandwidth. Researchers estimate that world-wide, organisations could collectively save over $1 billion a year in power costs by using such a technology.

The study group is essentially refiguring the process of auto-negotiation -- a link-detection technology in Ethernet, where a switch and NIC determine what speeds are supported (10/1000/1000Mbit/s) and establish the link rate. EEE would make this a more real-time process on Ethernet networks. For instance, a Gigabit-enabled laptop would switch to 10Mbit/s when idle, maybe 100Mbit/s during low-bandwidth activities, such as email or web surfing, and burst to 1000Mbit/s when downloading large files or streaming video.

"There's lots to take on with this effort," says Mike Bennett, senior network engineer at Lawrence Berkeley National Lab, and chair of the EEE Study Group.

Changing gear

One challenge is finding a way to make a PC or laptop network interface card (NIC) change gears more quickly -- "a couple orders of magnitude faster than auto-negotiation, to make the switch as seamless as possible," Bennett says. "Auto-negotiation runs at about one to 4 seconds and we're talking about -- just to start the discussion -- a millisecond of switching time."

EEE technology will have to work on both ends of a link to be successful, Bennett says. "When one device signals a speed change to another, the device would have to stop transmitting frames and tell the other end of the link, 'Hey, we're going to do a speed change here.'" The challenge with that is there are standard buffering sizes for Ethernet gear, he adds.

"Vendors build devices differently. Some have lots of buffers, some don't," he says.

If the IEEE and equipment vendors can figure all of this out, the savings could be huge for large organisations with thousands of Ethernet ports in PCs, servers and other devices, Bennett says.

Presentations given at EEE Study Group meetings cite a 2002 Department of Energy study estimating that the total power consumption of enterprise IT equipment in US offices at around 97 Terawatt hours per year, which translates to around $8 billion a year in energy costs. Extrapolating that cost over time, and accounting for network-related power consumption, the IEEE came up with the estimate of $450 million per year for the US. Bennett suggests it could be the same again each for Europe and the rest of the world.

"If all Ethernet ports in the US were suddenly EEE ports, you figure there's enough energy savings there at least worth thinking about," he says.

"We don't want to make it ridiculous and blow it up to something that isn't true, but those are reasonable estimates. If not, we would never have enough people interested in it to get a study group."

Defining the mechanisms

Discussions about how EEE technology will operate are in the early stages, Bennett says. What has emerged are two general concepts of what needs to happen at the Ethernet link layer, and higher layers. The first consideration is the control mechanisms that change the physical layer connection or PHY, to the desired speeds. This will define how to physically change PHY speeds among 10Base-T, 100Base-T, 1000T and 10GBase-T in that millisecond-changeover time period Bennett mentioned.

"First question is what's the mechanism of how you switch" among PHY speeds, says Ken Christensen, associate professor of computer science and engineering at the University of South Florida.

Christensen was one of the thinkers behind the idea of EEE, along with Bruce Nordman, a researcher at the Environmental Energy Technologies division of the US Lawrence Berkeley National Laboratory. Nordman has been researching the overall energy consumption of enterprise IT equipment, while Christensen has outlined a technology called Rapid PHY Selection (RPS). RPS involves ways LAN clients and switches can synchronise speeds through a quick MAC frame handshake. The trick is to handle variables such as bursty traffic.

"The next question is when do you switch, and that's the control policy," Christensen says.

Most of the IEEE's standardisation effort will focus on how to control PHY speed changes. The policies on when the changes occur will most likely be defined by makers of network equipment vendors, Christensen says. Vendors who figure out how to do this well could use that as a competitive advantage.

"Very simplistically, you could look at buffer thresholds, perhaps," to determine when to switch PHY speeds, he adds. "Or you could make an explicit measurement of utilisation. If it's more than a threshold, for example, you go up or down, in terms of link rate, which determines how much power the NIC draws from the system."