Degree

Doctor of Philosophy (PhD)

Department

Electrical and Computer Engineering

Document Type

Dissertation

Abstract

The compiler is able to detect the data dependencies in an application and is able to analyze the specific sections of code for parallelization potential. However, all of these techniques provided by a compiler are usually applied at compile time, so they rely on static analysis, which is insufficient for achieving maximum parallelism and desired application scalability. These compiler techniques should consider both the static information gathered at compile time and dynamic analysis captured at runtime about the system to generate a safe parallel application. On the other hand, runtime information is often speculative. Solely relying on it doesn't guarantee maximal parallel performance. So collecting information at compile time could significantly improve the runtime techniques performance.

The goal is achieved in this research by introducing new techniques proposed for both compiler and runtime system that enable them to contribute with each other and utilize both static and dynamic analysis information to maximize application parallel performance. In the proposed framework, a compiler can implement dynamic runtime methods in its parallelization optimizations and a runtime system can apply static information in its parallelization methods implementation. The proposed techniques are able to use high-level programming abstractions and machine learning to relieve the programmer of difficult and tedious decisions that can significantly affect program behavior and performance.

Date

8-3-2017

Committee Chair

Ramanujam, Jagannathan

Share

COinS