Selecting Memory Layouts for Model-Based Microarchitectural Performance Evaluation
Abstract
Virtual memory research increasingly relies on empirical performance models that use simulated address translation metrics (x)---such as TLB miss rate and page-walk cycles---to predict benchmark runtimes (y) for new hardware designs. The accuracy of these models depends on the training data: a set of (x,y) pairs collected by executing the benchmark on real hardware with memory layouts that mix pages of different sizes. Because the number of possible layouts grows exponentially with address space size, a key challenge is to select layouts that yield informative training samples spanning a broad range of x values. Previously proposed selectors rely on flawed heuristics that can leave large gaps in the model's input domain, necessitating a costly manual search for additional layouts. We solve this problem with Moselect, a feedback-driven layout selector that uses previously observed x values to steer selections toward uncovered regions. Moselect eliminates the manual effort, achieves good x domain coverage, converges faster than prior selectors, and significantly reduces worst-case runtime prediction error. Moreover, the higher-quality training data Moselect yields allow simpler simulators and runtime models to achieve worst-case prediction errors similar to those of their more sophisticated counterparts.