Power Capping To Peak Power Consumption - HP Integrity NonStop NB54000c Introduction Manual

Continuing with this example, Figure 11 shows the Insight Control Group Power Consumption graph for the group
of eight servers running a typical variable load application. Insight Control generates this graph using data
collected from each server's power management system. The power management system measures power
consumption twice per second and records both the peak and average power consumption. For peak power
consumption, each point in the graph represents the highest half-second power measurement recorded during a
given 5-minute period. The average power consumption denotes the arithmetic mean of all half-second power
measurements recorded over the same timeframe.
Figure 11. Group Power Consumption graph for a group of eight ProLiant DL380 G5 servers
The peak power consumption for the server group running this particular workload is about 3116 watts. The
average power consumption is about 1900 watts. Knowing the history of peak and average power consumption for
the group provides a good starting point for deciding how to set effective power caps that meet different power
management objectives in the data center.

Power capping to peak power consumption

As the power consumption graph in Figure 11 shows, peak power consumption for the eight-server group running
this workload is consistently about 3116 watts. Setting a power cap at this level would have no affect on overall
server performance. A 3116-watt power cap will ensure that the server group's power consumption will not
normally exceed 3116 watts for any significant period. For basic Power Capping, an administrator could safely
define cooling requirements for the server group against 3116 watts of power consumed rather than against the
larger maximum power number of 3384 watts that the HP Power Calculator Utility shows for this configuration. A
guardband (described later in this paper) is not necessary in this example for two reasons. First, eight servers are
enough for the average meter error to be small. Second, in this example the caps are set to the measured peak,
17