面向大数据的闪存三维堆叠存储结构研究

发布时间：2018-03-31 04:17

本文选题：大数据　切入点：三维堆叠　出处：《国防科学技术大学》2014年硕士论文

【摘要】：目前,大数据应用发展迅速,其具有与传统应用不同的数据特征,这给处理器的访存性能提出了挑战,加大了计算性能和存储性能之间的差异。三维堆叠存储结构以其低延迟、大容量、高带宽等特点,有效缓解了存储墙的问题,非易失存储介质除非易失性之外在容量和能耗上也优势明显,其与三维堆叠技术结合形成的非易失三维堆叠存储结构已成为发展的必然。闪存作为目前最成熟的非易失存储介质,更具有实际意义。本文据此重点研究针对大数据应用的闪存三维堆叠存储结构。首先,本文根据三维堆叠存储结构的现状,归纳出针对存储层次、存储器类型、处理器层和存储器层之间的通信方式和面向大数据应用这四个方面的三维堆叠存储结构的重点研究方向;主要分析了两种不同存储介质的三维堆叠存储结构:DRAM(动态随机存取存储器)堆叠和NVM(非易失性存储器)堆叠,据此总结了三维堆叠存储结构的特点,并针对大数据应用指出未来可能的研究思路。针对非易失三维堆叠存储结构,本文重点介绍了闪存三维堆叠存储结构,同时简要介绍新型非易失堆叠存储结构,并以此作为研究基础。其次,利用模拟实现和测试的方法,本文对闪存三维堆叠存储结构的特性进行了分析,其中选取能反映出闪存特性和其作为非易失存储介质的通性作为测试变量,以访存延迟作为结果进行测试。通过对多个测试数据的分析总结,最终得到四个典型特性,主要体现在三维堆叠结构的带宽优势在大数据应用上影响显著;随着带宽优势的增加,闪存的读写和擦除延迟影响逐渐减少;闪存页大小主要影响命中率,对延迟结果影响并不明显;闪存容量对结果影响较大。同时也给出了下一步的存储优化点。之后,根据由于闪存和DRAM两者不同的特性反映出的存储优化点,为了优化原有的闪存堆叠存储结构,充分利用两种不同存储介质的特性,本文提出了两个优化策略,分别是粒度差异优化策略和闪存层读写分区优化策略,旨在进一步提高系统访存性能,并据此给出了两个策略的设计思路和具体实现等理论基础。最后,本文对两个策略的优化效果进行了测试。通过对已有代码进行修改和利用工具收集必要的信息,本文实现了一个能反映闪存三维堆叠存储结构模块功能的简单模拟器,通过模拟器和选取合适的实验参数,得到多个结果,实验结果表明,两个策略优化了原有闪存三维堆叠存储结构下的处理器的访存性能,并且在闪存层读写分区策略下,闪存的写耐性也有一定的提升。在访存延迟结果下,粒度差异策略提高性能达11.3%,读写分区策略达5.4%,两个策略共同作用下,优化效果显著,性能提高了17.2%。综上,本课题针对大数据应用特征,主要对闪存三维堆叠存储结构进行了特性分析,并进行了优化,对进一步研究具有重要借鉴意义。
[Abstract]:At present, the rapid development of large data applications, the data has different characteristics from traditional applications, this to the processor memory performance is challenged, increasing the difference between computing performance and storage performance. The 3D stacked memory structure with its large capacity, low latency, high bandwidth characteristics, effectively alleviate the memory wall problem nonvolatile storage medium, unless the volatile in capacity and energy consumption have obvious advantages, non-volatile storage three-dimensional stacked structure has become the inevitable development of the combination with 3D stacked technology. Flash memory as the most mature nonvolatile storage medium, has more practical significance. This paper focuses on the application of data flash three-dimensional stack storage structure. First of all, according to the present situation of the 3D stacked memory structure, summed up for the memory hierarchy, the type of memory, the processor between layer and layer through memory The key research direction of 3D stacked memory structure and channel way for big data applications of these four aspects; the main analysis of the three-dimensional stack storage structure with two different storage medium: DRAM (dynamic random access memory (NVM) and stacked non-volatile memory) stack, then summarizes the characteristics of 3D stack storage structure. According to the application of big data pointed out possible future research ideas. The nonvolatile three-dimensional stack storage structure, this paper introduces the 3D stacked flash storage structure, at the same time, this paper introduces a new type of nonvolatile storage stack structure, and as the basis of the research. Secondly, using the simulation method to implement and test, the characteristics of 3D flash the stack storage structure is analyzed, which can reflect the characteristics and selection of flash memory as non continuity nonvolatile storage medium is used as test variables, the memory latency as a node The fruit was tested. Through the analysis of multiple test data summary, finally got four typical characteristics, mainly reflected in the three-dimensional structure of bandwidth advantage in large data applications significantly affected; with the increase of bandwidth advantages, flash memory read and write and erase the effects of delayed decreasing; flash memory page size mainly influences the hit rate. The results of the delay effect is not obvious; the capacity of flash memory have great influence on the results. At the same time, the storage optimization of the next point is given. Then, according to the optimal storage characteristics of flash memory and DRAM both reflect the different, in order to optimize the structure of the original flash storage stack, making full use of characteristics of two kinds of storage medium. This paper puts forward two optimization strategies are granularity optimization strategy and flash layer read-write partition optimization strategy, to further improve the performance of memory system, and gives two strategies The design and implementation of the theoretical basis of this paper. Finally, the optimization results of two strategies were tested. Based on the existing code to modify and use tools to collect the necessary information, this paper implements a simple simulator can reflect the three-dimensional flash storage structure module function, select the suitable experimental parameters by the simulator and and get multiple results, the experimental results show that the two strategies to optimize the original three-dimensional structure of the flash storage processor memory performance, and in a flash memory read and write layer partition strategy, flash write resistance has a certain increase in memory latency results, the size difference strategy to improve performance by up to 11.3% read and write, partitioning strategy 5.4%, two interaction strategy, the optimization effect is obvious, to improve the performance of 17.2%. in this subject according to the application characteristics of big data, mainly for flash 3D stack The characteristics of the storage structure are analyzed and optimized, which is of great significance for further research.

【学位授予单位】：国防科学技术大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP333

【相似文献】