Copyright by William Lloyd Bircher 2010 - The Laboratory for ...

can be considered as predictors which always predict the next phase to be the same as the
last. This approach works well if the possible transition frequency up the adaptation is
greater than the phase transition frequency of workload. Also, the cost of each transition
must be low considering the frequency of transitions. In real systems, these requirements
cannot currently be met. Therefore, the use of power adaptations does reduce
performance to varying degrees depending on workload. The cost of mispredicting
performance demand is summarized below.
Underestimate: Setting performance capacity lower than the optimal value causes
reduced performance. Setting performance capacity lower than the optimal value may
cause increased energy consumption due to increased runtime. It is most pronounced
when the processing element has effective idle power reduction.
Overestimate: Setting performance capacity higher than the optimal value reduces
efficiency as execution time is not reduced yet power consumption is increased. This
case is common in memory-bound workloads.
Optimization Points: The optimal configuration may be different depending on which
characteristic is being optimized. For example, Energy×Delay may have a different
optimal point compared to Energy×Delay 2 .
6.1.3 Performance Effects
P-states and C-states impact performance in two ways: Indirect and Direct. Indirect
performance effects are due to the interaction between active and idle cores. In the case
105