Last week, I installed a new 2.6.25.4 kernel on CoMO’s HPC cluster. The cluster consists of 9 IBM System x3550 machines with each 2 quad core Xeon E5320 CPUs (1.8 Ghz) and 8 GB RAM connected by a gigabit ethernet switch. According to the PDU, under full load, the power consumption is about 12.1 A (which should equal about 2800 Watt). Now in the new kernel I enabled some power saving features, such as cpufreq (ondemand governor), tickless kernel (dynamic ticks), USB suspend, etc. When the cluster is idle, it only uses 9.4 A now.. That’s a reduction by more than 20%! I still don’t understand why not all distributions are activating the ondemand governor on all systems by default in their init scripts… The power usage under full load does not seem to have much changed at first sight, which is not very surprising.
When the master node and the gigabit switch were turned on, but the 8 compute node powered off, power usage was 2.8 A. After cutting of the power to the 8 compute nodes on the PDU, power usage was reduced to 1.4 A. So 8 servers which are switched off but still connected to the power, use almost as much as one server which is powered on!
You should have a look at theses interesting readings :
http://mjg59.livejournal.com/88608.html
http://mjg59.livejournal.com/88951.html
http://mjg59.livejournal.com/89111.html
Interesting opinion, but I think this guy is strongly exaggerating, and the numbers above prove that CPU frequency scaling is useful and makes a difference. Of course it makes no sense to use CPU frequency scaling if CPU intensive tasks are running, but in that case, the CPU will simply go back to the normals peed thanks to the ondemand governor.