Many investigations reduce the simulation costs for DSE by analyzing program characteristics.
Hamerly et al. [2006] simulated only some representative program phases
rather than the whole program. From the perspective of program, Joshi et al. [2006]
found a reduced representative subset of programs by cluster analysis based on inherent
microarchitecture-independent characteristics. Moreover, statistical simulation
was employed to construct a synthesized shorter program to emulate the execution
characteristics of the original program [Genbrugge and Eeckhout 2009]. Unlike the
above approaches, predictive modeling techniques reduce simulated design configurations
by learning the relationship between design parameters and processor responses.
Following the supervised learning framework, the preceding task was accomplished by
linear regression model [Joseph et al. 2006] or Artificial Neural Networks (ANNs) [¨Ipek
et al. 2006; Khan et al. 2007; Cho et al. 2007; Dubach et al. 2011], where ANNs
are most widely used. Inspired by active learning, ¨Ipek et al. [2006] proposed the
intelligent sampling technique to enhance the accuracy of the supervised ANN
approach. This technique repeatedly updates an ensemble of 10 ANNs trained
by 10-fold cross validation over the labeled design configurations, and iteratively
labels (simulates) the unlabeled configurations on which the ANNs present largest
disagreements. Benefited from these techniques, architects no longer need to simulate
an excessively large number of design configurations. However, since the usefulness
of unlabeled design configurations is ignored, the above approaches still suffer from
either high simulation costs (for achieving high accuracies) or low prediction accuracy
(given limited computational resources).