面向动态众核的遥感数据处理任务双向调度方法研究

For massive remote sensing product, the mixed features with sequence and parallel had proposed high demands on computing resource dynamic allocation. Architecture based on dynamic Many Integrated Core (MIC) is of great value to solving computing resource variability in the process of computing-intensive remote sensing product processing tasks. Towards remote sensing product processing based on dynamic MIC parallel computing architecture, there exists system-level scheduling and task-level scheduling. It had been a research focus that requirements variability of task-level scheduling onto computing resource could not reach to system-level core scheduling timely. Logical core fusion algorithm towards dynamic MIC and task tree construction based on mixed features of remote sensing product would be further studied. Remote sensing parallel processing queue is designed upon revised dynamic backfill algorithm so that computing resource distribution transformation of task-level scheduling can give feedback to system-level scheduling. And dynamic split of logical core and fusion of basic core would be come true. The proposed work in this proposal is beneficial to promote the application of dynamic MIC in massive remote sensing product parallel processing and scheduling.

海量遥感数据产品的串/并行混合处理对计算资源的动态分配提出了更高需求，动态众核架构对于解决计算密集型遥感数据处理任务中的计算需求变化问题具有重要研究意义与应用价值。基于动态众核架构的遥感数据处理需要综合考虑系统级的逻辑计算核心动态调度与任务级的数据处理流程调度，如何解决任务级调度对计算资源的需求变化无法及时反馈给系统级的核调度，难以获得相匹配的计算资源问题，是遥感数据处理领域计算密集型应用的研究难点。本项目开展动态众核逻辑核聚合(Fuse)与解耦(Split)算法研究、遥感数据混合处理任务树构建、动态资源模型优化等方法的研究，基于动态回填算法实现双向调度模型的优化，使任务级调度的资源变化需求能够及时映射到系统级的核调度，实现逻辑核的动态解耦与基本核的聚合，使计算资源利用效率得到大幅提升。本项目的开展将有助于为数据密集型遥感数据并行处理与调度技术研究提供新的方法和思路。

本项目针对动态众核构建的海量遥感数据并行计算环境中，现有数据处理流程调度算法无法充分利用动态众核逻辑核动态可变特性，造成计算资源的极大浪费问题，提出了基于 Backfill资源模型的 Backward-Forward 双向调度算法，使动态众核的系统级调度与遥感数据处理流程的任务级调度相匹配。并基于 CoreFusion众核架构，研究了通过聚合(Fuse)与解耦(Split)进行 Forward 调度队列动态更新，实现海量遥感处理过程中的并行资源高效调度，为众核技术在遥感数据并行处理领域的应用与推广奠定理论与实践基础。