伴随大数据量的应用任务在中央处理器 (CPU)与图形处理器(GPU)组成的异构处理平台上的部署日益广泛, 如何高效利用 GPU硬件中的并行资源, 成为亟待解决的问题。通过对单 GPU任务映射策略进行研究, 提出多 Stream有向无环图 (MS-DAG)任务映射策略。通过分析 DAG图中的节点依赖关系, 根据节点依赖关系的不同, 划分合理的并行分支, 利用多 Stream流水线并行的方式, 实现适合GPU硬件特点的任务映射策略。通过与HEFT在不同条件下的性能对比, 可以看出: 当HEFT算法中的各处理器性能不一致时, MS-DAG任务映射策略的任务映射效率相比 HEFT算法有约 10%的提升; 当 HEFT算法中的各处理器性能一致时, MS-DAG任务映射策略的任务映射效率相比 HEFT算法有30%的提升。

Utilizing Graphic Processing Unit(GPU) parallel resource efficiently as the application task of large data quantity deployed on the rapid development CPU+GPU heterogeneous platform became an urgent problem. A task mapping strategy of Multiple Stream-Direction Acyclic Graph(MS-DAG) is proposed, based on the research of task mapping strategy of single GPU. The task mapping strategy for GPU hardware is realized by analyzing the node dependencies in DAG graph, according to the difference of node dependency, dividing reasonable parallel branches and using multi stream pipelining. It shows that the task mapping efficiency of MS-DAG task mapping strategy is about 10% higher than that of Heterogeneous Earliest Finish Time(HEFT) algorithm when the performance of each processor is inconsistent in HEFT algorithm; and the task mapping efficiency of MS-DAG task mapping strategy is 30% higher than that of HEFT algorithm, when the performance of each processor in the HEFT algorithm is consistent.