Dynamic time management for improved accuracy and speed in host-compiled multi-core platform models

With increasing complexity and software content, modern embedded platforms employ a heterogeneous mix of multi-core processors along with hardware accelerators in order to provide high performance in limited power budgets. Due to complex interactions and highly dynamic behavior, static analysis of real-time performance and other constraints is challenging. As an alternative, full-system simulations have been widely accepted by designers. With traditional approaches being either slow or inaccurate, so-called host-compiled simulators have recently emerged as a solution for rapid evaluation of complete systems at early design stages. In such approaches, a faster simulation is achieved by natively executing application code at the source level, abstracting execution behavior of target platforms, and thus increasing simulation granularity. However, most existing host-compiled simulators often focus on application behavior only while neglecting effects of hardware/software interactions and associated speed and accuracy tradeoffs in platform modeling. In this dissertation, we focus on host-compiled operating system (OS) and processor modeling techniques, and we introduce novel dynamic timing model management approaches that efficiently improve both accuracy and speed of such models via automatically calibrating the simulation granularity. The contributions of this dissertation are twofold: We first establish an infrastructure for efficient host-compiled multi-core platform simulation by developing (a) abstract models of both real-time OSs and processors that replicate timing-accurate hardware/software interactions and enable full-system co-simulation, and (b) quantitative and analytical studies of host-compiled simulation principles to analyze error bounds and investigate possible improvements. Building on this infrastructure, we further propose specific techniques for improving accuracy and speed tradeoffs in host-compiled simulation by developing (c) an automatic timing granularity adjustment technique based on dynamically observing system state to control the simulation, (d) an out-of-order cache hierarchy modeling approach to efficiently reorder memory access behavior in the presence of temporal decoupling, and (e) a synchronized timing model to align platform threads to run efficiently in parallel simulation. Results as applied to industrial-strength platforms confirm that by providing careful abstractions and dynamic timing management, our models can achieve full-system simulations at equivalent speeds of more than a thousand MIPS with less than 3% timing error. Coupled with the capability to easily adjust simulation parameters and configurations, this demonstrates the benefits of our platform models for early application development and exploration.