SnoopDog: memory overflow watchdog based on snooper

Abstract

This thesis proposes a new memory error detection approach that is based on snooper hardware that observes memory traffic flowing through the system bus. Memory error detectors have been mainlly built using software instrumentation, which suffer from performance and compatibility issues. Memory snooper-based approach can transparently monitor the memory access behavior of the a process. How- ever, to observe a software from the memory’s point of view, we need to overcome microarchitectural barriers such as cache. This thesis explores what are the architectural and software level constraints in the snooper-based detector and then implemented a redzone-based heap memory bug detector.