PPTM:一种面向异构系统的主动式任务映射方法

PPTM:a proactive task mapping method for heterogeneous systems

在数据高速增长的背景下,异构计算作为满足新兴应用不断提高的算力需求的有效途径,涌现了许多异构加速系统。在这些异构加速系统中,高效的任务映射是充分发挥加速器潜能提升应用程序性能的关键之一。先前工作提出了许多基于有向无环图如何最小化应用程序整体执行时间和最小化异构多处理器之间通信开销等高效的任务映射方法,这些工作通常采用将任务映射到加速器上来提高整个应用的性能。但某些应用程序如果将所有子任务全部映射到加速器上执行,会带来额外的通信开销,进而可能达不到提升性能的预期,甚至造成整个应用程序的性能下降。因此,本文提出了一种基于预测的主动式任务映射算法(PPTM)来应对这样的场景,实现高效的任务映射。实验表明,本文算法能够更准确感知计算任务的运行时状态,大幅提高应用程序的整体性能。

With the rapid growth of data,heterogeneous computing has emerged as an effective way to meet the ever-increasing computing power demands of emerging applications,and many heterogeneous acceleration systems have emerged.In these heterogeneous acceleration systems,efficient task mapping is one of the keys to making full use of the accelerator’s potential to improve application performance.In the previous work,many efficient task mapping methods have emerged based on directed acyclic graphs to minimize the overall execution time of the application and minimize the communication overhead between heterogeneous multiprocessors.In these work,it is usually assumed that mapping tasks to accelerators will definitely bring performance benefits to the entire application.However,it was found that if some applications map all subtasks to the accelerator for execution,it will bring additional communication overhead,which may not meet expectations for performance improvements,and even cause the performance of the entire application to decrease.Therefore,a prediction-based proactive task mapping algorithm(PPTM) is proposed to deal with such scenarios and achieve efficient task mapping.Experimental results show that the prediction algorithm can more accurately perceive the runtime state of computing tasks,and then through the mapping algorithm,the whole performance of the application can be greatly improved.