Exploring the Potential of Performance Monitoring Hardware to Support Run-time Optimization

Date The final copy of this thesis has been examined by the signatories, and we find that both the content and the form meet acceptable presentation standards of scholarly work in the above mentioned discipline. Run-time optimization defines the process of dynamically modifying an applica-tion's characteristics to promote desirable execution behavior. Since there is a wealth of information available at runtime which is unavailable to static compiler analysis, run-time optimization has substantially more potential to fully utilize processor resources. A critical component of run-time optimization systems is the run-time profiler which must accurately capture specific aspects of application execution behavior while maintaining a low overhead. Unfortunately, most existing profiling approaches cannot meet these constraints and therefore cannot feasibly be deployed in a run-time optimization system. While modern microprocessors can collect run-time information through on-chip Hardware Performance Monitoring (HPM) support, it is not clear whether this technology can effectively guide a run-time optimization framework. To date the HPM information of various processor systems has almost solely been used in post-execution performance tools. This thesis evaluates the potential of performance monitoring hardware to support profiling for run-time optimization. The trade-offs in meeting the constraints imposed in a run-time environment are analyzed by evaluating various sampling rates and analysis techniques. Altogether, the thesis characterizes the amount of information available through PMU sampling as well as the extent in which compiler analysis can extend PMU information. Path profiling and code coverage analysis, important elements of run-time optimization, are evaluated to demonstrate the effectiveness of run-time profiling with hardware support. Dedication To my father, mother, and brother. v Acknowledgements First and foremost, I would like to thank my advisor Dan Connors for his guidance in this work. With his deep insights, generous advice, and continuous encouragement, I have learned and accomplished more than I could have imagined when I first began graduate school. I look forward to continuing my research with him as I pursue a Ph.D degree. I would like to acknowledge Andrew Pleszkun and Manish Vachharajani, members of my defense committee, for their valuable advice and suggestions. I express my thanks to the entire DRACO research group. Matthew Iyer was instrumental to the development of ideas as well as the implementation of the work in this thesis. Alex Settle mentored me early on and was extremely helpful in teaching me research tools and skills. Vijay Janapa Reddi provided me with a lot of …