A Machine Learning and Compiler-based Approach to Automatically Parallelize Serial Programs Using OpenMP

Single core designs and architectures have reached their limits due to heat and power walls. In order to continue to increase hardware performance, hardware industries have moved forward to multi-core designs and implementations which introduces a new paradigm in parallel computing. As a result, software programmers must be able to explicitly write or produce parallel programs to fully exploit the potential computing power of parallel processing in the underlying multi-core architectures. Since the hardware solution directly exposes parallelism to software designers, different approaches have been investigated to help the programmers to implement software parallelism at different levels. One of the approaches is to dynamically parallelize serial programs at the binary level. Another approach is to use automatic parallelizing compilers. Yet another common approach is to manually insert parallel directives into serial codes to achieve parallelism. This writing project presents a machine learning and compiler-based approach to design and implement a system to automatically parallelize serial C programs via OpenMP directives. The system is able to learn and analyze source code parallelization mechanisms from a training set containing pre-parallelized programs with OpenMP constructs. It then automatically applies the knowledge learned onto serial programs to achieve parallelism. This automatic parallelizing approach can be used to target certain common parallel constructs or directives, and its results when combined with a manual parallelizing technique can achieve maximum or better parallelism in complex serial programs. Furthermore, the approach can also be used as part of compiler design to help improve both the speed and performance of a parallel compiler