Statistical inference for efficient microarchitectural analysis

The transition to multiprocessors expands the space of viable core designs and requires sophisticated optimization over multiple design metrics. However, microarchitectural design space exploration is often inefficient and ad hoc due to the significant computational costs of hardware simulators. Long simulation times cause designers to subjectively constrain the design space considered. However, by pruning the design space with intuition before a study, the designer risks obtaining conclusions that simply reinforce prior intuition, thereby limiting the study's value. Addressing these fundamental challenges in microarchitectural analysis becomes increasingly urgent as the semiconductor industry moves into new domains where tried and tested intuition becomes less effective. This dissertation presents the case for statistical inference in microarchitectural design, proposing a simulation paradigm that (1) defines a comprehensive design space, (2) simulates sparse samples from that space, and (3) derives inferential regression models to reveal salient trends. These regression models accurately capture performance and power associations for comprehensive multi-billion point design spaces. Moreover, they are capable of thousand's of predictions per second. Used as computationally efficient surrogates for detailed simulation, regression models enable previously intractable analyses of performance and power. Leveraging model efficiency, this dissertation demonstrates qualitatively new capabilities by using pareto frontiers to identify power-efficient designs, contour maps to visualize bottlenecks, and roughness metrics to quantify non-monotonicity in design topologies. Furthermore, inferential models enable qualitatively new capabilities in optimization for emerging design priorities. Not only do these models answer prior questions far more quickly, they answer new questions previously intractable with detailed simulation. This dissertation implements robust optimization techniques to assess multiprocessor heterogeneity and microarchitectural adaptivity, quantifying trends and limits in performance and power efficiency from these design paradigms. The capabilities from inference scale to multi-billion point design spaces, giving designers the holistic view necessary to successfully implement the transition to multiprocessors.