Data-Oriented Runtime Scheduling Framework on Multi-GPUs

摘要

GPU has been generally accepted as an efficient accelerator in the field of high performance computing (HPC). On some heterogeneous systems, multiple GPUs are installed on each computing node. To make things more complicated, these GPUs may even have different architectures. Therefore, it is a challenge to efficiently schedule tasks and data on heterogeneous system. In this paper, we present DoSFoG, a data-oriented runtime scheduling framework on heterogeneous system equipped with multiple GPUs. In DoSFoG, the data blocks, instead of tasks, are taken as the scheduling units. It uses a data-oriented directed acyclic graph (DoDAG) as representation of an application, which is proved to be equivalence to task DAG. Based on DoDAG, a runtime scheduling framework is designed. Besides, a hierarchical storage structure is carefully designed based on the various levels of memory in the system. Page-locked memory and soft cache on GPU device memory are used to improve the data transfer. DoSFoG is evaluated with different applications on a system equipped with different GPUs. The results show that DoSFoG can achieve high data locality, scalability, load balance and performance improvement for large size of data.