基于混合内存的系统级数据划分策略研究

发布时间：2018-06-21 16:33

本文选题：相变存储器 + 混合内存　；参考：《杭州电子科技大学》2017年硕士论文

【摘要】：为了解决日益增长的主存能耗问题,以及满足越来越高的主存容量需求,一种新型存储技术相变存储器(PCM,Phase Phase-Chage Memory)应运而生。PCM具有字节可寻址,高存储密度和低能耗的诸多优势,成为了动态随机存取存储器(DRAM,Dynamic Random Access Memory)的有力的替代技术。但是由于PCM的读写不对称性以及有限的写入寿命,PCM目前尚无法完全取代DRAM主存的位置。于是,同时使用两种存储技术的混合内存架构,是当前主要的研究方向。对于将DRAM和PCM作为同级主存的混合内存架构来说,面临的最大挑战就是如何划分数据。页面划分策略会直接影响到系统的平均访存时间和PCM材料的寿命。目前,对于这种混合内存架构页面划分策略的研究,普遍采用先对页面进行简单划分,当页面表现出不符合预期的读写时,再将页面通过迁移的方式迁至不同的存储器中,来实现对读写操作分布的控制。基于迁移的划分策略没有考虑迁移操作本身的高成本以及迁移操作带来的额外写操作,这些迁移操作在很大程度上延长到了系统的平均访存时间并带来了额外的能量开销。首先,本文介绍了混合内存页面划分策略的研究背景、研究意义以及国内外研究现状,并介绍了所涉及的混合架构及仿真器的相关技术。其次,本文详细分析了现有混合内存划分策略的优点和缺点,并根据现有划分策略具有问题,提出了基于虚拟页访存行为的混合内存页面划分策略(Virtual Page Behavior based Page Management Policy for Hybrid Main Memory,VPBM)。通过模拟器在云端获取页面的访存行为数据,根据页面的访存行为数据来预测虚拟页的访存行为,对页面直接进行划分。因为相同进程虚拟页访存行为在多次运行中是相似的,通过历史的访存行为数据来分配页面,可以避免程序运行过程中进行页面迁移,提升系统效率和降低访存能耗,延长PCM寿命。然后,本文针对多核多进程并发的环境下VBPM获取访存行为难度几何倍数升高的问题,提出了一种通过单进程访存行为预测多进程环境下访存行为自适应划分策略。通过对页面访问频率的预测,来决定多进程环境下页面划分与抢占的策略。在仿真器获取的访存行为数据足够多的前提下,根据页面已访问次数和预期访问次数的差值,可以计算出访问的偏差情况,根据预期访问频率计算后的结果与竞争页面预期频率的比较结果,来决定页面的最终位置。这种自适应划分策略有效降低了多任务并发时模拟器仿真成本,提高了系统的整体效率。最后,本文对所做工作与重要贡献做出了总结,并对未来的工作方向做出了展望和规划。
[Abstract]:In order to solve the problem of increasing main memory energy consumption and to meet the increasing demand of main memory capacity, a new storage technology, PCMN Phase Phase-Chage memory (PCM), emerges as the times require. PCM has the advantages of byte addressable, high storage density and low energy consumption. It has become a powerful alternative to dynamic Random Access memory (DRAM). However, due to the asymmetry of reading and writing in PCM and the limited write lifetime, PCM can not completely replace the main memory of DRAM. Therefore, the hybrid memory architecture with two storage technologies is the main research direction. For the hybrid memory architecture with DRAM and PCM as the same level main memory, the biggest challenge is how to divide the data. Page partitioning strategy will directly affect the average access time of the system and the lifetime of PCM materials. At present, for the research of page partitioning strategy of hybrid memory architecture, it is generally used to divide the pages simply first, and then move the pages to different memory by migration when the pages do not meet the expectations of reading and writing. To control the distribution of read and write operations. The partition strategy based on migration does not take into account the high cost of the migration operation itself and the additional write operation brought by the migration operation, which to a large extent extends to the average memory access time of the system and brings additional energy cost. Firstly, this paper introduces the research background, significance and research status of hybrid memory page partitioning strategy, and introduces the related technologies of hybrid architecture and simulator. Secondly, the advantages and disadvantages of the existing hybrid memory partitioning strategy are analyzed in detail, and according to the problems of the existing partitioning strategy, a virtual page behavior based hybrid memory page partitioning strategy is proposed. Through the simulator in the cloud to obtain the page access behavior data, according to the page access behavior data to predict the virtual page access behavior, the page is divided directly. Because the virtual page access behavior of the same process is similar in many operations, it can avoid the page migration, improve the system efficiency and reduce the energy consumption by distributing the page through the historical memory access behavior data. The life of PCM was prolonged. Then, in order to solve the problem of increasing the geometric multiple of VBPM access behavior in multi-core and multi-process concurrent environment, an adaptive partition strategy is proposed to predict the memory access behavior in multi-process environment by single-process memory access behavior. The strategy of page partitioning and preemption in multi-process environment is determined by predicting the frequency of page access. On the premise that the emulator acquires enough memory access behavior data, the deviation of access can be calculated according to the difference between the number of page visits and the expected number of visits. The final location of the page is determined by comparing the expected access frequency with the expected frequency of the competing page. This adaptive partition strategy can effectively reduce the simulation cost of multi-task concurrent simulator and improve the overall efficiency of the system. Finally, this paper summarizes the work done and important contributions, and makes a prospect and plan for the future work direction.
【学位授予单位】：杭州电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP333

【参考文献】