项目名称: 异构多核平台上基于软件分布式共享内存的编程模型研究
项目编号: No.61202049
项目类型: 青年科学基金项目
立项/批准年度: 2013
项目学科: 计算机科学学科
项目作者: 李波
作者单位: 浙江工业大学
项目金额: 23万元
中文摘要: 异构多核是由架构不同的核心组成的计算平台,由于其能够在特定的领域中针对应用的特点充分发挥不同处理器核的作用,成为了高性能计算平台主流架构和硬件基础。但也由于其不同架构组合复杂多样,传统的单一消息或者共享的编程模型都难以清晰描述各个计算部件之间的协同关系,导致其面临着编程要求高、产能低和调试的困难等挑战。本项目在由GPU集群构成的异构多核计算平台上,以软件分布式共享内存模型为基础,研究基于GPU设备内存的分布式共享设备内存的编程模型,达到简化编程,优化性能的目的。重点研究如下几个问题:(1)基于GPU设备内存、主存的两级共享内存的数据一致性维护;(2)GPU设备内存地址空间到主存地址空间的映射机制;(3)CPU-GPU间负载合理分配策略;(4)共享内存中数据预取策略。本项目旨在能为降低异构多核平台上的编程复杂性,提高编程效率奠定理论和技术基础。
中文关键词: 高性能计算;GPU;编程模型;性能优化;
英文摘要: Heterogeneous multicore architecture has been becoming the mainstream architecture in high performance computing community. However, due to its architecture is so complexity and diversity, neither of the traditional mainstream message-passing and shared memory programming models could describe the role of the computing units. So the some challenge is unavoidable to face, include programing on it is very difficult, achieving productivity and portability is also hard. This project aims to meet these challenges through studying a new programming model based on software distributed shared memory. The research focus on the following topics :(1) Data consistency maintenance through main memory to the GPU device memory.(2) the address space mapping mechanism from GPU device memory to main memory.(3) the load balancing strategy between CPU and GPU. (4) the data prefetching scheme for shared data. The project aims to do some fundamental research to reduce the programming complexity for the heterogeneous multicore platforms and improve programming productivity.
英文关键词: HPC;GPU;Programming model;Performance optimization;