Turn off your servers!

I was interested to see that research into power reduction is intensifying, to the point where serious research is going into how to switch off servers that are unnecessary at any one point in time. As ever, though, I suspect the problem will eventually boil down to good systems management.

Google recently made a research award of $1 million to a US-based research group, consisting of a collaboration of computer scientists from the universities of California, Michigan, Virginia, and Rutgers University.

They're researching ways of redesigning server processors so that power is routed separately to the CPU core and to the memory controller. This will, in theory at least, allow the memory controller to remain alive, presumably keeping the contents of RAM refreshed, while the power-hungry core transistors are asleep. This means the CPU can power back up quickly when needed.

The award was part of the the first round of the company’s new Google Focused Research Awards, which is worth $5.7 million all told.

Turning servers off is getting easier anyway as a result of the increased focus by component and server vendors on power management, so the Google award is a potential step further in the same direction.

The driver behind the trend is of course that it's very expensive to keep hundreds of servers running. Fortunately, one of the advantages of virtualisation is that server images can easily be moved around the datacentre and concentrated, especially at off-peak times, so that they occupy a minimal number of hosts. Provided you can ensure that each host has the right mix of servers so that no host is, for example, saddled with a dozen I/O-intensive applications, virtual image mobility can save significant amounts of energy, if you can manage -- or even bring yourself -- to turn off those that aren't doing anything (I can see the dialog box now: are you really, really sure?).

However, the number of management applications that have the smarts to understand and manage the huge range of technologies involved is pretty small. So what's key now is to find software that can keep all the plates spinning. Let us know when you find it...