一种针对NAND Flash的缓存管理算法研究

发布时间：2018-05-16 15:29

本文选题：固态硬盘 + NAND　；参考：《国防科学技术大学》2012年硕士论文

【摘要】：基于NAND flash的固态硬盘在访问速度、功耗和抗震性等方面比机械硬盘具有更多优势。但由于擦块写页、擦写次数有限和读写擦代价不对称性等特点，使得其在小数据随机写情况下性能会明显下降。特别是在多通道并发架构下，过多的擦除操作严重威胁着固态硬盘的性能和寿命，缓存管理对于改善这方面的性能具有重要作用。本文根据现有FAB、CLC等块级缓存置换算法和页填充策略，分析了其基本思路和优缺点。对缓存管理算法中的缓存置换算法和页填充策略分别进行了改进，，结合NAND Flash的操作特点和当前多通道并发架构的要求，设计了一种新的缓存管理算法。本文的主要研究成果如下：（1）在CLC置换算法的基础上，设计了基于动态驻留优先级的缓存置换算法。不同于通过固定的链表长度实现数据驻留优先级的变化，本文以最近访问距离为阈值判定热数据是否应降为冷数据。围绕该思想，设计了与之相适应的页簇状态、双链表维护、置换对象的选择等内容。仿真和测试证明，该改进使得对数据“价值”的评价更加准确，能够提高命中率，减少擦除次数。（2）在单阈值的页填充策略的基础上，设计了一种基于双阈值的页填充策略。不同于单阈值的页填充策略，根据部分页填充和全页填充对实时性影响的不同，设定双阈值判定二者执行的时机。基于这一思想，对部分页填充和全页填充分别在缓存环节和垃圾回收环节的时间开销进行了分析，推导出合适的阈值表达式，并通过仿真给出了阈值的具体值。仿真和测试证明，该改进能够进一步提高吞吐量。最后将基于动态驻留优先级的缓存置换算法和基于双阈值的页填充策略整合为一个缓存管理算法，并在基于多通道并发架构的硬件平台上进行了性能测试。测试结果表明该算法在小数据随机写情况下能够改善系统性能。
[Abstract]:Solid state hard disk based on NAND flash has more advantages than mechanical hard disk in access speed, power consumption and seismic resistance. However, due to the characteristics of block writing, limited erasing times and asymmetry of the cost of reading and writing, the performance of small data random writing will be significantly decreased. Especially in multi-channel concurrent architecture, too much erasure is a serious threat to the performance and lifetime of solid-state hard disk. Cache management plays an important role in improving the performance of this aspect. Based on the existing block level cache replacement algorithms and page filling strategies, this paper analyzes their basic ideas, advantages and disadvantages. The cache replacement algorithm and the page-filling strategy in the cache management algorithm are improved respectively. A new cache management algorithm is designed according to the operation characteristics of NAND Flash and the requirements of the current multi-channel concurrent architecture. The main research results of this paper are as follows: 1) based on the CLC permutation algorithm, a cache permutation algorithm based on dynamic resident priority is designed. Different from the change of data dwell priority by a fixed length of linked list, this paper uses the nearest access distance as the threshold to determine whether the hot data should be reduced to cold data. According to this idea, the corresponding page cluster state, double linked list maintenance, replacement object selection and so on are designed. The simulation and test show that the improved method makes the evaluation of data "value" more accurate, and can improve the hit ratio and reduce the erasure times. 2) based on the single threshold page-filling strategy, a double-threshold page-filling strategy is designed. Different from the single threshold page-filling strategy, according to the effect of partial page-filling and full-page-filling on real-time performance, two thresholds are set to determine the timing of their execution. Based on this idea, the time cost of partial page filling and full page filling in cache and garbage collection are analyzed, and the appropriate threshold expression is deduced, and the specific value of threshold is given by simulation. Simulation and test show that the improvement can further improve the throughput. Finally, the cache replacement algorithm based on dynamic resident priority and the page-filling strategy based on double threshold are integrated into a cache management algorithm, and the performance is tested on the hardware platform based on multi-channel concurrency architecture. The test results show that the algorithm can improve the system performance in the case of small data random writing.
【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】：2012
【分类号】：TP333

【参考文献】