高并发数据访问模式的基础理论与系统设计

Memory system has become the most critical performance bottleneck of all computing systems, from multi-core processors to supercomputers. Meanwhile, more emerging data intensive applications and new memory technologies arise. It is urgently needed for the research community to establish the corresponding basic theory to overcome the new challenges. As Moore's law will end soon, this project proposes using architectural approach to solve the growing memory wall problem. Based on our preliminary studies, in the frontier of international research, we plan to build the basic theory of concurrent data accesses, and conduct corresponding system designs. Specially, we will conduct this research in four aspects: 1) Establishing high concurrent data access model and building the minimum complete set of metrics for data access patterns, and forming the periodic table of data access patterns that is similar with that in pysical world. 2)Building the relationship model between data access concurrency and memory system performance; 3) Proposing the optimization algorithm of data access patterns in hybrid memory system; 4) Proposing the optimization algorithm of remote data access patterns in distributed shared memory system. The research result of this project will answer a series of 12 basic scientific problems about data access patterns. Besides the completeness of theory that is important for correctness, we will keep the theory simple and easy to use to perform optimizations dynamically during runtime. The research result will play an important role in optimizing the performance of data intensive applications and optimize the efficiency of supercomputers. In the research project, at least six international top level papers will be published, and at least three US or Chinese patents will be applied.

存储系统已成为从多核处理器到超级计算机的各种规模计算机系统中最严重的性能瓶颈，各种新兴数据密集型应用与新型存储技术不断呈现，面对新的挑战，学术界亟需建立相应的基础理论。适应后摩尔定律时代的要求，本课题提出用体系结构的方法尤其是并发的方法解决日益严重的存储墙问题，在前期研究基础上，立足国际研究前沿，构建高并发数据访问的基础理论，具体包含四个方面的研究工作：1）建立高并发数据访问模式的数学模型和用于解决存储墙问题的“访存模式周期表”；2）建立数据访问并发度与数据访问性能之间的关系模型；3）提出混合存储系统中访存模式的优化算法；4）提出分布式共享存储系统中远地访存模式的优化算法。本课题将回答关于访存模式的12个基本的科学问题，注重理论的简洁性、易用性、完备性，将对优化数据密集型应用的性能、提高超级计算机系统效率发挥重要作用。本项目拟发表6篇以上国际顶级水平论文，并申请3项以上美国或中国发明专利。

存储系统已成为从多核处理器到超级计算机的各种规模计算机系统中最严重的性能瓶颈，各种新兴数据密集型应用与新型存储技术不断呈现，面对新的挑战，学术界亟需建立相应的基础理论。适应后摩尔定律时代的要求，本课题提出用体系结构的方法尤其是并发的方法解决日益严重的存储墙问题，在前期研究基础上，立足国际研究前沿，构建高并发数据访问的基础理论，超额完成了研究目标，取得成果20余项，其中高水平期刊论文5篇，高水平会议论文4篇，学术专著2部，发明专利3项，华为技术大会特邀报告1项，科技成果大规模转化与应用1项，培养硕士博士多名。