面向异构众核系统的非规则问题优化技术研究

项目名称： 面向异构众核系统的非规则问题优化技术研究

项目编号： No.61303050

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： 王珏

作者单位： 中国科学院计算机网络信息中心

项目金额： 27万元

中文摘要： 随着高性能计算应用和能耗需求的日益增长，以异构众核系统（CPU为主处理器，众核为协处理器）作为集群节点已经成为未来高性能计算机发展的一个重要趋势。该系统具有多存储层次、多编程模型、多并行计算模式等特点，传统的优化技术已经很难满足非规则问题所带来的存储访问不规则、输入数据敏感和负载不均衡等问题。这就需要重构非规则问题算法库，深度挖掘算法之间相关性、自身局部性、并行性和非规则等特点，充分发挥异构众核系统特点，提高程序性能。为此，本课题针对典型的非规则问题算法库（排序算法库和稀疏矩阵向量乘算法库）在异构众核节点（传统CPU + Intel MIC协处理器）研究动态负载均衡和数据管理优化的基础上选取调优参数，通过定量化和试验相结合的方式建立优化技术的性能模型,进行算法库自动调优研究。本项目的研究成果将很好地提高非规则问题算法库的生成和执行效率，提高程序员的生产效率，具有重要的理论意义和应用价值。

中文关键词： 并行计算；非规则问题；自动调优；性能模型；众核

英文摘要： With the requirement of HPC applications and energy, the node with hybrid many cores systems (CPU is host processor, many core is the co-processor) are getting more and more important for the development of high performance computer. The system characteristics consist of multiple level storages, multiple programming models, multiple parallel computing models, etc. Traditional optimization technologies are difficult to satisfy the irregular accessing, input-intensive, overhead imbalance problems. It is necessary to reconstruct the irregular problem algorithm libraries, and to find the dependencies of algorithms, self-locality, parallelism, and irregularity for exploring the characteristics of hybrid many cores systems. We focus on the scheduling of dynamic overhead, data management for irregular problem algorithm libraries (sorting library and sparse matrix vector multiply) on hybrid many core systems (CPU + Intel MIC). We then select the tuning parameters based on the optimization technologies. The performance model is built using quantification and trial methods. Based on this model, we build the auto-tuning algorithm library. The research results will improve the performance to build and execute the irregular problem library, and increase the productivity of programmers.

英文关键词： parallel computing；irregular problem；autotuner；performance model；many cores