Ethernet Goes Green

On July 20, new EPA specifications for computers come online; the agency estimates that if U.S. businesses purchased only Energy Star-compliant systems next year, they could save $1.2 billion over the lifetimes of their new computers.

And the effort isn't stopping there. The IEEE's Energy Efficient Ethernet (EEE) group is studying how to cut Ethernet's power requirements through a technology called RPS, or Rapid PHY Selection. The EEE says its effort could save $450 million per year in electricity costs.

Anyone who's watched the energy tab for Ethernet switches, routers, desktops and servers rise should welcome the opportunity to buy more efficient systems. It's a little too early to tell, however, if the news is all good.Despite promising developments, there are still big questions. Primary among these, for IT, is whether Ethernet can become more efficient without additional costs--in terms of IT support or productivity, or the expense of adding RPS to NICs--that outweigh the incremental savings. The biggest fear is that there will be incompatibilities similar to those that existed, and some say still exist, for autonegotiation, which is used to negotiate the speed and duplexity of twisted-pair connections automatically. These incompatibilities have slowed network installations while network engineers diagnosed the resultant connectivity problems and implemented manual workarounds, thus increasing support costs.

Continue Reading This Story...

Power-hungry Nics

The EEE initiative is built on two assumptions. First, the various speeds of twisted-pair Ethernet--10/100/ 1,000/10,000 Mbps--have consume power at different rates. The faster the link, the more power consumed, even while the link is not transmitting data. Second, most computers, especially desktops, are often idle with very short bursts of data usage. Initial studies by Ken Christensen, a researcher with the University of South Florida, confirm this. The EEE is trying to figure out how to adjust Ethernet speed and corresponding power usage to reflect the instantaneous bandwidth requirements. This involves watching for changes in network utilization or queue depths and quickly adjusting the link speed as needed. If a computer is simply idle, for example, or the user is working with applications that require low levels of bandwidth, a Gigabit NIC would power down to 100 Mbps.

The benefit of such a speed adjustment is that Ethernet NICs at the desktop or server, as well as at the switch, would draw less power, which translates into lower utility bills.The best way to gauge the potential benefits is to consider power consumption of Ethernet at different speeds. The biggest savings would be in a scenario in which a 10-Gbps Ethernet link reduces power consumption to the level of 1 Gbps when data is not transmitted. If you assume 10 Gbps uses about 10 watts for each of the two ends of the link and 1 Gbps uses about 2 watts per end, and the cost of electricity is 8 cents per kilowatt hour, you could save $5.60 per link per year if the computer is in power-saving mode for 50 percent of each day.

Under those conditions, a site with 1,000 computers would save $5,600 per year in electricity costs. There are also side benefits in cooling costs and switches may realize additional power savings as they reduce port power usage.

Technical DifficultiesThe EEE work will have to go through a number of standards review steps to address implementation issues and gain consensus from all the EEE-participating vendors. We're unlikely to see final approval before the end of 2009--and even then, you'll need EEE-capable equipment on both ends of an Ethernet link to benefit.

Meantime, a number of technical challenges must be worked out. The biggest is how to increase or decrease Ethernet speed quickly enough to avoid packet loss and the resultant problems with apps and the end-user experience. Developers from companies participating in the EEE are working to address this. Current tests have shown that the time it takes to make the required change in Ethernet speed can vary significantly, from a few microseconds to 1 to 20 milliseconds. Once you get into the latter range, there's potential for noticeable packet drops, enough to cause TCP to slow down retransmissions. One possible solution is to use some type of flow control during the speed change, but this could create its own problems if switch buffers fill up and have to drop packets anyway. And imagine trying to do this through a VoIP phone that's being used as a switch for a locally attached PC.

Exchanging Frames

You may be thinking that the EEE's methods sound a lot like autonegotiation. But autonegotiation is too slow to make RPS practical, though early research with RPS (known in 2004 as ALR, or Adaptive Link Rate) by Bruce Nordman of Lawrence Berkeley National Labs and South Florida's Christensen used autonegotiation to change speeds, to test the general concept.

RPS must use a faster mechanism than autonegotiation, something that probably will involve a simple frame exchange. Given the problems with autonegotiation, we hope initial versions of RPS work more smoothly. That won't be a problem, according to Howard Frazier, Broadcom's technical director for technical strategy, who is heavily involved with the EEE initiative and chaired the 802.3 group that standardized autonegotiation. He stressed that lessons learned from autonegotiation will be applied to RPS, and that implementing RPS is less risky than implementing autonegotiation because RPS can be turned off.We hope Frazier is right and, in fairness, the IEEE has plenty of opportunity to mitigate potential problems on this project. However, dynamically changing the speeds at both ends of a network connection in a multivendor network could have serious consequences if it goes awry. If, like autonegotiation, one end changes speed and the other end isn't notified of the change, the link will stop working. The feature will also have to be built into new NICs at a low-enough cost to reap the energy benefits.

Peter Morrissey, an NWC contributing editor, is manager of network design and development at Syracuse University and an adjunct professor at Syracuse&Apos;S School of Information Studies. Write to him at [email protected].