lammert - 3:58 am on Jul 15, 2010 (gmt 0)

The current cluster of servers I'm maintaining for webserving and other purposes mainly uses multicore Intel type CPUs. The servers have been bought or leased on different times for different purposes and there are therefore some different Intel processor series in use.

When buying a server, processing power, memory, disk space and price are often the reason to select a specific configuration. But with current energy prices and limitations in data centers on the power usage per server, it is also important to look at the average power usage. A not too special server uses about 150 Watt. With a price of $0.15 per kWh, that translates to $197 per year. Prices may vary in your situations, but it is a significant amount of money. For servers in data centers you should add to that the power consumption of the cooling units to pump the heat of that server away.

The main power user in most servers is the CPU. Modern multicore processors have the ability to automatically reduce the internal clock speed when the full processing power is not needed and this lowers power usage significantly. Most web servers don't need the CPU running on maximum speed continuously anyway. All Linux servers I have running use the Munin monitor software to make graphs of all kind of system parameters during runtime. It makes it easy to tune the system configuration and decide which resources are used to their max. Some months ago I wrote a plugin to monitor the internal clock frequency of the individual cores of the processors and there are huge differences between the XEON families, which I want to share here.

The low end server model uses a dual core XEON 3065 processor of 2.33GHz. On low load, this processor steps down to 2GHz. Not a real reduction in speed. More interesting to see is that at high load, both cores are always stepped up together to 2.33GHz. This processor doesn't seem to have the option to regulate the core frequencies independently.

The second server model I watched is based on the Intel 6420 core 2 duo series running at max 2.13GHz. This processor is able to reduce the CPU frequency to 1.60 GHz. It also switches the core frequencies independently when more processing power is needed.

The third server model runs the quad core XEON 3430 with a maximum frequency of 2.40GHz. With low CPU utilization, the CPU frequency is reduced to 1.20GHz. Just as with the 6420, this model is also capable of switching the frequency of each core independently.

These results show, that the maximum processing power a CPU can deliver is no direct measure for the electrical power it consumes during life time when the CPU is not constantly under high load. Newer more expensive CPU models with better frequency control may be cheaper in the long run. Also new fabrication techniques reduce average power consumption for newer CPU models with the same specs as older CPUs. Therefore, if power consumption and electricity bills are a concern for you, it might be wise to consider looking at the current CPUs in use and upgrade the bad performers.