While heterogeneous architectures are increasing popular with High Performance Computing systems, their effectiveness depends on how efficient the scheduler is at allocating workloads onto appropriate computing devices and how communication and computation can be overlapped. With different types of resources integrated into one system, the complexity of the scheduler correspondingly increases. Moreover, for applications with varying problem sizes on different heterogeneous resources, the optimal scheduling approach may vary accordingly. Thus, we introduce a Profile-based AI-assisted Dynamic Scheduling approach to dynamically and adaptively adjust workloads and efficiently utilize heterogeneous resources. It combines online scheduling, application profile information, hardware mathematical modeling and offline machine learning estimation modeling to implement automatic application-device-specific scheduling for heterogeneous architectures. A hardware mathematical model provides coarse-grain computing resource selection while the profile information and offline machine learning model estimates the performance of a fine-grain workload, and an online scheduling approach dynamically and adaptively distributes the workload. Our scheduling approach is tested on control-regular applications, 2D and 3D Stencil kernels (based on a Jacobi Algorithm), and a data-irregular application, Sparse Matrix-Vector Multiplication, in an event-driven runtime system. Experimental results show that PDAWL is either on-par or far outperforms whichever yields the best results (CPU or GPU).

虽然异构架构在高性能计算系统中越来越流行，但它们的有效性取决于调度程序在将工作负载分配到适当的计算设备上的效率以及通信和计算如何重叠。随着不同类型的资源集成到一个系统中，调度器的复杂度也相应增加。此外，对于在不同异构资源上具有不同问题大小的应用程序，最佳调度方法可能会相应变化。因此，我们引入了基于 Profile 的 AI 辅助动态调度方法来动态和自适应地调整工作负载并有效利用异构资源。它结合了在线调度、应用程序配置文件信息、硬件数学建模和离线机器学习估计建模，以实现异构架构的自动应用设备特定调度。硬件数学模型提供粗粒度计算资源选择，而配置文件信息和离线机器学习模型估计细粒度工作负载的性能，在线调度方法动态和自适应地分配工作负载。我们的调度方法在事件驱动的运行时系统中在控制规则应用程序、2D 和 3D Stencil 内核（基于 Jacobi 算法）和数据不规则应用程序、稀疏矩阵向量乘法上进行了测试。实验结果表明，PDAWL 要么达到标准，要么远远优于产生最佳结果（CPU 或 GPU）中的任何一个。