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.
3.5 Methodology <strong>for</strong> Power Modeling<br />
With an understanding of system level events that are visible to the processor it is<br />
possible to apply the iterative modeling process as depicted in Figure 3.5. This procedure<br />
utilizes linear and polynomial regression techniques to build power models <strong>for</strong> individual<br />
subsystems. <strong>The</strong> user identifies workloads which target a particular subsystem (cache,<br />
system memory, disk) and per<strong>for</strong>ms regression modeling using per<strong>for</strong>mance events as<br />
inputs. <strong>The</strong> model is then applied to a larger set of workloads to confirm accuracy and<br />
the lack of outlier cases. Depending on the outcome, the process is repeated with<br />
alternate per<strong>for</strong>mance events as inputs. Though an exhaustive search of per<strong>for</strong>mance<br />
events can be per<strong>for</strong>med, a rapid solution is found when events are selected with high<br />
correlation to subsystem activity. Details of the modeling process in Figure 3.5 are listed<br />
below.<br />
1. Measure subsystem-level power using subset of workloads. Begin with simple,<br />
easy-to-run workloads.<br />
2. Confirm that Coefficient of Variation is greater than α <strong>for</strong> the chosen workload.<br />
<strong>The</strong> simplest workloads often do not generate sufficient power variation <strong>for</strong> model<br />
tuning. For example consider any of the cache-resident workloads in SPEC CPU 2000<br />
which generate little or no activity in subsystems outside of the processor cores such as<br />
memory. Tuning the model based on these low-variation workloads may cause the<br />
process to include per<strong>for</strong>mance events that do not correlate well with power.<br />
40