Low-overhead Online Code Transformations.

Abstract

The ability to perform online code transformations - to dynamically change the implementation of running native programs - has been shown to be useful in domains as diverse as optimization, security, debugging, resilience and portability. However, conventional techniques for performing online code transformations carry significant runtime overhead, limiting their applicability for performance-sensitive applications. This dissertation proposes and investigates a novel low-overhead online code transformation technique that works by running the dynamic compiler asynchronously and in parallel to the running program. As a consequence, this technique allows programs to execute with the online code transformation capability at near-native speed, unlocking a host of additional opportunities that can take advantage of the ability to re-visit compilation choices as the program runs.

This dissertation builds on the low-overhead online code transformation mechanism, describing three novel runtime systems that represent in best-in-class solutions to three challenging problems facing modern computer scientists. First, I leverage online code transformations to significantly increase the utilization of multicore datacenter servers by dynamically managing program cache contention. Compared to state-of-the-art prior work that mitigate contention by throttling application execution, the proposed technique achieves a 1.3-1.5x improvement in application performance. Second, I build a technique to automatically configure and parameterize approximate computing techniques for each program input. This technique results in the ability to configure approximate computing to achieve an average performance improvement of 10.2x while maintaining 90% result accuracy, which significantly improves over oracle versions of prior techniques. Third, I build an operating system designed to secure running applications from dynamic return oriented programming attacks by efficiently, transparently and continuously re-randomizing the code of running programs. The technique is able to re-randomize program code at a frequency of 300ms with an average overhead of 9%, a frequency fast enough to resist state-of-the-art return oriented programming attacks based on memory disclosures and side channels.