多核系统上分层MPT模型的联合调度和映射

Joint scheduling and mapping of hierarchical MPT model on multi-core systems

为了实现异构多核系统上的任务的并行执行,提出了一种分层多处理器任务模型(MPT)的联合调度和映射策略。首先,基于分层多处理器任务模型,考虑异构多核平台上混合数据和任务并行以及进程之间的依赖关系,建立起一个由多个MPT构成的MPT图;然后,基于MPT图提出了一种分层调度算法,调度算法将图结构划分为若干层,每层采用符号核组的一个集合,每个符号核组可依次执行若干个MPT;为了实现一个MPT程序到一个多核系统的感知映射,又提出了3种从符号核到物理核的映射,以实现多处理器并行任务的计算。实验结果表明,提出的分层MPT模型的联合调度和映射方案相比于其他方案不仅可以降低执行时间,而且还可以提高加速性能。

In order to realize parallel execution of tasks on heterogeneous multi-core systems, a joint scheduling and mapping strategy for hierarchical multi-processor tasks model is proposed in this paper. First, based on the hierarchical multi-processor tasks(MPTs) model, a MPT graph composed of multiple MPTs is established by considering the parallelism of mixed data and tasks and the dependencies between processes on heterogeneous multi-core platform. Then, a hierarchical scheduling algorithm based on MPT graph is proposed. The scheduling algorithm divides the graph structure into several layers. Each layer adopts a set of symbol core groups, and each symbol core group can execute several MPTs in turn. In order to realize the perceptual mapping from an MPT program to a multi-core system, three kinds of mappings from symbolic core to physical core are proposed to implement the computation of multi-processor parallel tasks. Experimental results show that the proposed hierarchical MPT model can not only reduce the execution time, but also improve the acceleration performance compared with other schemes.