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.
per<strong>for</strong>mance demand, reactive power management increases per<strong>for</strong>mance capacity<br />
sometime after the phase transition. During the time between the change in demand and<br />
capacity, per<strong>for</strong>mance may be less than optimal. Similarly, power consumption is sub-<br />
optimal on transitions from high to low demand. <strong>The</strong> amount of per<strong>for</strong>mance loss is<br />
proportional to the number of phase changes in the workload and the lag between demand<br />
and capacity. For increasing per<strong>for</strong>mance in power-limited situations, reactions must be<br />
fast to prevent overshooting the power limit or missing opportunities to increase<br />
per<strong>for</strong>mance.<br />
Identifying when to adapt is complicated <strong>by</strong> the presence of multiple cores sharing power<br />
resources. Consider Figure 7.1. Core-level power consumption is shown <strong>for</strong> a system<br />
with multiple simultaneous threads. <strong>The</strong> program threads are fixed to the cores with<br />
thread N on core N, thread N-1 on core N-1, etc. Since power monitoring is typically<br />
provided at the system-level [Po10], existing power control techniques use the erratic<br />
fluctuations in the total power <strong>for</strong> predicting future behavior. This is un<strong>for</strong>tunate since in<br />
this example, the individual threads have a periodic, easily discernable pattern, while the<br />
pattern in the aggregate power is less discernable. If power phases can be tracked at the<br />
core-level, accurate dynamic power management schemes can be devised.<br />
To improve the effectiveness of power management the use of predictive, core-level<br />
power management is proposed. Rather than reacting to changes in per<strong>for</strong>mance<br />
demand, past activity patterns are used to predict future transitions. Rather than using<br />
aggregate power in<strong>for</strong>mation, activity and power measured at the core-level is used.<br />
122