Genetically Programmed Response Surfaces for Efficient Design Space Exploration
چکیده
In spite of many efforts to speed up cycle-accurate architecture simulation, exponential increases in architectural design complexity threaten to make traditional design optimization techniques completely intractable. Response surface methodologies address this challenge by transforming the optimization process from a lengthy series of detailed simulations into the tractable formulation and rapid evaluation of a marginally less accurate but easy to evaluate analytical expression—a predictive model. We propose genetic programming as a powerful method for creating these predictive response surface models out of sampled architectural performance data. Genetically programmed response surfaces (GPRSs) allow the architect to make rapid design optimizations (because only a small number of detailed simulations are needed) while simultaneously obtaining insight into the problem domain (because the resulting response surface — a non–linear polynomial in our case — exposes relationships and relative weights among the design variables). We validate our methodology on realistic datasets and compare it to recently proposed techniques for predictive design space exploration. GPRSs are highly accurate when making global predictions about architectural performance behavior based on only small samples of performance data: global predictions of IPC incur less than 3% mean percentage error based on sample sizes of less than 1% of one target processor design space, and no worse than than mean 6% error at sample sizes as small as 0.0000002% out of over one billion possible design points from a second target space. GPRSs can therefore reduce required simulation costs by up to six orders of magnitude.