Data centers are often designed (or overdesigned) to handle the highest load that might come their way, which really doesn't jibe with the green ideal.

Part of the problem is the fickle nature of CPU and chipset power saving techniques we tested (see main story) may not translate into linear power savings because of where the power company measures consumption for billing. Uninterruptible power supplies (UPSs) electrically work like large capacitors in that they deliver power in a linear fashion, backfilling electrons smoothly as they are consumed. CPU conservation efforts designed to save server power aren't realized where power is metered by the power company in the same linear fashion that they are served up to the UPS. The UPS generally stores enough energy for all items plugged into it plus extra capacity needed in case of a service interruption from the power company. This means the draw of the UPS on the power company's transformers isn't a direct reflection of the power that in-line devices are using.

Secondly, the process of cooling CPUs and other server components is often geared toward the highest cooling requirement. Network operations center (NOC) temperatures are often measured in five minute (or longer) periodic samples, and cooling is adjusted according to the rise and fall of ambient temperature (or the average of either single or multiple area sensors). Even though a single server's power consumption is likely to vary widely (in our tests, as much as 17 per cent), cooling must be available for the maximum heat output mandated by the server's requirements.

As servers in racks may also be affected by other server temperature rises and chills within the server or 'temperature zone', servers tend to average each other's ambient temperature through active cooling (fans and heat dissipation design). This makes it more difficult still to react to the fast, tiny changes in CPU and chipset temperatures as they wake-up and sleep for just milliseconds at a time.

The bottom line is NOCs don't react to power consumption and heat dissipation in direct correlation to a server's needs, and design practices tend to sway settings to the highest consumption/chilling needed, rather than reacting in complete harmony with rapid power consumption characteristics dictated by green principles.

There is some hope, though. UPS and NOC design vendors are slowly reacting to the need for green sensing, allowing power distribution units (PDU) to actively sense power consumption and report power usage as well as detecting when servers seem to be unused so they can be powered down completely until needed. So-called 'smart' PDUs are now available from vendors such as Schneider Electric/APC and Raritan that have monitoring software to detect power consumption within threshold settings. VMware ships virtual machine management software that detects machines with no loads and shuts them down or 'spins them up' based on policies set by administrators.