Adaptive Task Granularity Strategy for OpenMP3.0 Task Model on Cell Architecture

Considering that the best strategy which controls task granularity usually depends on the application characteristics, we propose an adaptive task granularity (AdaptiveTG) strategy, which first applies breadth-first policy until all the available threads are busy. And then if some thread becomes idle, the work stealing technique, which provokes the victim to create a task with sufficient size at the oldest spawnable point, is introduced. The strategy doesn’t only maximize the task granularity but also efficiently solves the unbalanced computation loads. The experiment is conducted on Cell architecture on a set of kernels based on task model. Evaluations indicate that the adaptiveTG technique achieves speeds factors from 4.1 to 7.2 compared to the serial executions, and outperforms Tascell framework and AdaptiveTC strategy for most applications.