Scalable Emulation of Heterogeneous Systems

The breakdown of Dennard's transistor scaling has driven computing systems toward application-specific accelerators, which can provide orders-of-magnitude improvements in performance and energy efficiency over general-purpose processors. To enable the radical departures from conventional approaches that heterogeneous systems entail, research infrastructure must be able to model processors, memory and accelerators, as well as system-level changes---such as operating system or instruction set architecture (ISA) innovations---that might be needed to realize the accelerators' potential. Unfortunately, existing simulation tools that can support such system-level research are limited by the lack of fast, scalable machine emulators to drive execution. To fill this need, in this dissertation we first present a novel machine emulator design based on dynamic binary translation that makes the following improvements over the state of the art: it scales on multicore hosts while remaining memory efficient, correctly handles cross-ISA differences in atomic instruction semantics, leverages the host floating point (FP) unit to speed up FP emulation without sacrificing correctness, and can be efficiently instrumented to---among other possible uses---drive the execution of a full-system, cross-ISA simulator with support for accelerators. We then demonstrate the utility of machine emulation for studying heterogeneous systems by leveraging it to make two additional contributions. First, we quantify the trade-offs in different coupling models for on-chip accelerators. Second, we present a technique to reuse the private memories of on-chip accelerators when they are otherwise inactive to expand the system's last-level cache, thereby reducing the opportunity cost of the accelerators' integration.