Tales from the Engine Shed

When I got my new laptop (an HP Pavilion DV6262 if you are interested, and yes I am happy with it so far) it took me a while to get Ubuntu fully working on it and so for a short while I was using Windows Vista. When I got Ubuntu working it appeared that battery life was significantly shorter than under Windows Vista and although I now attribute that to the gnome power monitor being more pessimistic than that of Windows (and having a few bugs, like the most erratic estimation of remaining time I have ever seen) at the time I was looking for ways to stretch the battery life when running Linux.

A quick bit of research suggested I should try CPU frequency scaling which slows the CPU down when demand on it’s resources are lowered. This was relatively easy and didn’t appear to affect the functioning of the system much. I didn’t conduct any test and just took it for read that I was saving energy. A few weeks later and I was discussing strategies for saving energy in data centres since that is now becoming a large factor in the running costs, and hit upon the idea of using CPU frequency scaling to reduce the thirst for energy a densely populated rack has. This sat at the back of my mind for a while but never got much further.

This week I decided to conduct some experiments using my desktop PC, a Dell Optiplex GX280. This is a fairly stable and powerful machine and whatever my opinion of Dell server and laptop offerings (low) I think they present extremely good value for money in the domestic PC market. This machine has a 2.8GHz Pentium 4 processor and with CPU frequency scaling enabled and the governor set to ‘On Demand’ I regularly see the CPU running at 350MHz - surely there is a saving to be had here?

It seems not. When running in ‘Performance’ mode (2.8Ghz all the way) my average power draw measured over a 24 hour period is 118.1 watts. For the same period running in ‘On Demand’ mode I was drawing 117.4 watts - hardly a significant saving. I know the frequency scaling works as if I choose ‘Powersave’ mode (350 MHz all the way) my load average climbs to around 6 and Firefox becomes unusable. My only conclusion is that the power supply carries on drawing the same amount of power regardless of the demand of the system. If this is true for all power supplies then I think there is a very real need for hardware manufacturers to develop more intelligent power supplies. It might be that server-spec gear already does this, so maybe I shouldn’t give up completely on this project and see what the results are like for an HP Proliant server (my rack-mount hardware of choice). However at the minute maybe I should explore other avenues for minimising data centre power draw - virtualisation and clustering anyone?