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 ...
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
the case of I/O, the observed workloads make little or no use of disk and network<br />
resources. For the chipset subsystem, the causes are not as clear and require further<br />
study. As mentioned in the previous section the lack of disk power management causes<br />
little variation in disk power consumption. If these subsystems are to benefit from<br />
dynamic adaptation, workloads with larger variation in utilization would be needed.<br />
In order to justify the use of CoV <strong>for</strong> identifying workloads with distinct, adaptable<br />
phases the probability distributions <strong>for</strong> extreme cases are considered. In order <strong>for</strong> a<br />
subsystem-workload pair to be a strong candidate <strong>for</strong> optimization, it must have distinct<br />
program/power phases. If a workload exhibits constant power consumption it is difficult<br />
to identify distinct phases. Further if the difference in phases is small, it may be difficult<br />
to distinguish a phase in the presence of sampling noise. <strong>The</strong>re<strong>for</strong>e, a strong candidate<br />
should have multiple distinct phases. This can be observed in the power amplitude<br />
distributions in Figure 4.7. A narrow distribution represents a constant power<br />
consumption. A wide distribution represents many levels of power consumption.<br />
Distinct power phases only exist in the wide distributions.<br />
Figure 4.7 shows the average power distribution <strong>for</strong> the various subsystems. Not<br />
surprisingly, CPUs are the dominant power users. However, unlike distributed, scientific<br />
NAS Parallel Benchmark [FeGe05-1] and mobile, productivity workloads such as<br />
PCMark and 3DMark [MaVa04], I/O and disk power are significant. While differences<br />
in average subsystem power are large at 138% <strong>for</strong> disk compared to CPU, the variations<br />
within an individual subsystem are even greater. A comparison of subsystem power<br />
55