面向动态规划计算的并行编程模型和运行时系统研究

项目名称： 面向动态规划计算的并行编程模型和运行时系统研究

项目编号： No.61303021

项目类型： 青年科学基金项目

立项/批准年度： 2014

项目学科： 自动化技术、计算机技术

项目作者： 于策

作者单位： 天津大学

项目金额： 25万元

中文摘要： 动态规划广泛应用于科学和工程计算，问题规模增加以及并行计算硬件技术发展使动态规划计算并行程序研发的难度随之增高，本课题通过面向动态规划的可复用并行编程模型和高效运行时系统来简化此类应用并行编程的难度和复杂度。基于DAG数据驱动编程模型，研究自适应多层次任务调度方法，通过任务粒度二次划分以及多层次任务调度的方法解决负载不平衡问题；研究适用于不同硬件平台以及应用类型的大规模中间结果数据的处理方法，以高效共享文件的方式解决中间结果数据的存储和访问问题，以按需数据传输的方式减少不必要的通信；研发插件式DAG模式库管理方法，支持模式的复用和系统扩展。研发过程结合计算生物信息学和计算实验金融领域的具体算法和应用进行，以实际应用反馈研究方案，并针对集群、MPP体系结构以及具体文件系统进行优化。课题预期成果可为基于动态规划计算的应用研发提供编程模型和运行时系统的支持，缩短科学及应用研发周期。

中文关键词： 并行编程模型；动态规划；任务调度；数据存取；并行计算

英文摘要： Dynamic programming is widely used in science and engineering computing. The increasing scale of problem and development of parallel computing hardware made the parallel programming for dynamic programming get more difficult. We will decrease the difficulty and complexity of such parallel programming by reusable parallel programming model and high efficiency runtime system for dynamic programming. The self-adaptive multilevel task scheduling method is based on the DAG data driven programming model, to ensure the load balance by refining the scale of subtasks and multilevel scheduling. According to the large-scale intermediate result data for different hardware platform and specific applications, we propose the data storage and accessing methods based on shared files. And the on demand data transferring method will reduce the unnecessary communications. The DAG pattern library is based on plug-in architecture to support the reuse of pattern and the expansibility of the system. The research work will cooperate with the specific algorithms and applications from computational bioinformatics and computational finance to get the feed back from real usage. Optimization will be made according to Cluster and MPP architecture and several file systems. The expected result of this proposal will support the dynamic programmi

英文关键词： Parallel Programming；Dynamic Programming；Task Scheduling；Data Storage and Retrieving；Parallel Computing