Program Parallelizing Apparatus for Multithreading Optimization

The patent segments the fork point determination process into two distinct phases: (1) a rounding phase that quickly removes obviously suboptimal fork points based on static boost value thresholds, and (2) an optimization phase that refines the combination of remaining candidate fork points. This segmentation allows the system to handle large numbers of candidate fork points efficiently by eliminating low-value candidates early without sacrificing the quality of the final parallel execution performance.

The patent performs preliminary filtering of fork points by calculating static boost values and removing those below predetermined thresholds before proceeding to the optimization phase. This preliminary action reduces the search space for optimal fork point combinations, enabling faster determination of the optimal set while maintaining high parallel execution performance. The rounding step prepares the data structure in advance to facilitate more efficient optimization.

The patent introduces dynamic boost values that adapt based on the specific input data and execution context, moving away from static predetermined thresholds. The system calculates dynamic boost values by analyzing the actual execution trace, which allows the fork point determination to adapt to different input scenarios. This dynamic approach improves parallel execution performance while the rounding step with static thresholds provides a simple initial filtering mechanism, creating a balanced two-phase process.

The patent incorporates feedback mechanisms where the optimization phase uses information from the rounding phase and adjusts the fork point combination based on actual execution performance. The system iteratively refines the fork point selection by analyzing execution traces and feedback from the rounding step, allowing the process to learn from actual performance data and improve subsequent parallelization decisions.