Copyright by William Lloyd Bircher 2010 - The Laboratory for ...
Copyright by William Lloyd Bircher 2010 - The Laboratory for ...
Copyright by William Lloyd Bircher 2010 - The Laboratory for ...
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
5.3 Complete-System Desktop Power Model<br />
In this section results <strong>for</strong> the application of the trickle-down modeling approach are<br />
presented <strong>for</strong> a recent desktop plat<strong>for</strong>m. This plat<strong>for</strong>m differs from the previous server in<br />
terms of process technology, system architecture, manufacturer and workload among<br />
others. It is shown that though this plat<strong>for</strong>m is significantly different than the server, the<br />
trickle-down modeling approach still accurately models power. Of particular importance<br />
are two major differences: subsystem level power management and workload<br />
characteristics. Power management increases the complexity and utility of the power<br />
model as power consumption varies greatly with the application of power management.<br />
Compare this to the server system in which power remains near a constant level due to<br />
subsystems not reducing per<strong>for</strong>mance capacity, and there<strong>for</strong>e power consumption, during<br />
periods of low utilization. Increased power variation is also attributable to desktop-<br />
specific workloads. While server workloads tend to always operate at full speed (e.g.<br />
SPEC CPU) desktop workloads such as SYSmark and 3DMark contain large portions of<br />
low utilization. This exposes the impact of power management and the need to model it.<br />
5.3.1 System Description<br />
To validate the effectiveness of the trickle-down approach the process is applied to a<br />
recent desktop plat<strong>for</strong>m. A comparison of the two systems used in this study (server and<br />
desktop) is provided in Table 5.3. <strong>The</strong>se systems differ in their power management<br />
implementations and subsystem components. <strong>The</strong> desktop system is optimized <strong>for</strong> power<br />
84