Collaborative Research: A Study and implementation of Semantic Constructs for Highly Scalable Leading-edge Scientific Computing

将科学问题扩展到下一代HEC系统的10,000,000个处理器，今天受到编程模型、语言及其支持的编译系统的传统实践的严峻挑战。为了实现这一目标，人们必须暴露出比传统方法(如MPI)更大程度的并行度并提高并行计算效率。这个合作研究项目的目标是通过应用一种创新的编程模型来显著提高挑战性物理问题的可扩展性。该策略将分布式存储空间中使用全局屏障同步的静态消息传递过程粒度进程替换为使用分区全局地址空间中的轻量级同步对象的动态消息驱动多线程的模型。并行性是直接从大型、不规则、稀疏和时变的数据结构中提取的。短暂的用户线程将允许在数据结构上同时执行多个任务，从而暴露出内在的近细粒度并行性。通过异步消息驱动处理的高级通信策略，通过重叠计算和通信来隐藏系统范围内的延迟。因此，这将使一类目前不能用传统方法完成的物理问题成为可能。

Scaling scientific problems to 10,000,000 processors for the next generation HEC systems is today severely challenged by conventional practices of programming models, languages, and their supporting compilation systems. To achieve this goal one must expose greater degree of parallelism and improve parallel computing efficiency than is otherwise feasible with conventional methods such as MPI. The goal of this collaborative research project is to dramatically enhance the scalability of challenging physics problems, through the application of an innovative programming model. The strategy is to replace static message-passing course grained processes using global barrier synchronization in a distributed memory space with a model using dynamic message-driven multiple threads using lightweight synchronization objects in a partitioned global address space. Parallelism is to be extracted directly from the large irregular sparse and time varying data structures. Ephemeral user-threads will permit many simultaneous tasks over the data structures, exposing the intrinsic near-fine grain parallelism. System-wide latency will be hidden by overlapping computation with communication through the advanced communication strategy of asynchronous message-driven processing. Consequently this will enable a class of physics problems that cannot currently be done using conventional methods.