混合页缓存系统的性能建模与分析

发布时间：2018-01-15 22:16

本文关键词：混合页缓存系统的性能建模与分析　出处：《中国科学技术大学》2017年硕士论文　论文类型：学位论文

【摘要】：近年来,高性能计算和大数据分析的发展对页缓存系统的容量要求越来越高,然而动态随机存储器(Dynamic Random Access Memory,DRAM)由于价格昂贵和扩展性差等问题限制了页缓存的大规模扩充,所以一些研究人员提出结合非易失性存储介质(Non-volatile Memory.NVM)来构建大规模的混合页缓存系统。但是,从经济和技术的角度来看,NVM的加入是否能够提高页缓存系统的性能仍是一个值得探讨的问题。回答这个问题的关键是,结合NVM和DRAM在访问性能和价格方面的差异来考虑如何设计混合页缓存系统。本文提出了混合页缓存系统的分析模型,从而分析不同的设计选择对混合页缓存系统的性能影响。我们主要考虑了以下几个设计选择:(1)不同的架构选择——平行架构和分层架构,(2)不同的缓存替换策略选择,以及(3)不同的存储设备容量配置。为了分析混合页缓存系统的性能,我们刻画了页缓存系统的访问请求模型,然后使用马尔科夫过程来描述在不同的设计选择下的页缓存状态分布的变化并推导出了稳定时的状态分布,同时为了解决计算效率问题,我们使用平均场理论来近似对稳定时状态分布求解,并分别通过理论分析和实验验证证实了近似过程的可靠性。随后为了评估混合页缓存的性能,我们定义了一个基于访问延迟的页缓存性能评判标准。基于此模型,我们通过数值分析学习了如何合理地设计页缓存系统,以实现在页缓存系统的"高吞吐量"和"高灵活性"之间做一个良好地权衡。特别地,我们选择PCM作为非易失存储介质的一个实例进行数值分析,分析结果表明,即使在当前的PCM的性价比下,页缓存系统中加入PCM介质已经可以提高系统性能,且选择"平行架构"更有利于对性能的优化。而随着PCM技术的发展,特别是写性能得到飞速提升时,分层架构会表现出明显的优势。
[Abstract]:In recent years, the development of high performance computing and large data analysis requires the capacity of page cache system more and more. Dynamic Random Access Memory, however. DRAM) limits the large-scale expansion of page caching due to problems such as high prices and poor scalability. So some researchers proposed a combination of non-volatile memory. NVM-based non-volatile storage media to build large-scale hybrid page caching systems. Whether the addition of NVM can improve the performance of the page caching system from an economic and technical point of view is still a question worth discussing. The key to answer this question is. Considering the differences between NVM and DRAM in terms of access performance and price, we consider how to design a hybrid page cache system. So we analyze the impact of different design choices on the performance of hybrid page caching system. We mainly consider the following design options: parallel architecture and hierarchical architecture. In order to analyze the performance of the hybrid page cache system, we describe the access request model of the page cache system. Then the Markov process is used to describe the change of the state distribution of the page cache under different design choices, and the stable state distribution is deduced to solve the problem of computational efficiency. We use the mean field theory to approximate the steady state distribution, and verify the reliability of the approximate process by theoretical analysis and experimental verification, and then to evaluate the performance of mixed page cache. We define a page cache performance evaluation standard based on access delay. Based on this model, we learn how to design the page cache system reasonably by numerical analysis. In order to achieve a good trade-off between "high throughput" and "high flexibility" of a page cache system. In particular, we chose PCM as an example of a non-volatile storage medium for numerical analysis. The analysis results show that even under the current PCM performance-price ratio, the addition of PCM media to the page buffer system can improve the performance of the system. With the development of PCM technology, especially with the rapid improvement of write performance, hierarchical architecture will show obvious advantages.
【学位授予单位】：中国科学技术大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP333

【参考文献】