【作者】 高鹏；

【导师】 汪东升；

【作者基本信息】 清华大学 ， 计算机科学与技术， 2018， 博士

【摘要】 目前的计算机系统中,需要采用多种异质存储设备共同构成存储层次结构,才能满足处理器对存储器的延迟、吞吐量、容量、非易失性的需求。但是,这种架构会给软硬件设计带来更高的复杂性和功耗。而相变存储器的出现,为简化存储层次结构、构建未来存储设备带来了新的可能性。这是因为该技术几乎具有理想存储设备应有的所有特性。然而,相变存储技术也存在写入寿命较短的缺陷。因此,本文主要针对相变存储器的写入寿命问题展开了研究。由于相变存储器不仅对写入位置的重复程度敏感,而且对写入的内容同样敏感,本文首先研究了数据值局部性,这一描述数据内容分布的规律,并定量化的分析了数据值局部性的特征,然后从减少单次写入时的翻转位数和平衡写入位置间的写入量差异的角度展开寿命优化工作。本文主要成果及创新点如下:·本文详细分析了数据值局部性,提出了数据值局部性的宏观表现和微观表现的概念。对X86-32,X86-64以及arm-32下的各类应用的所使用到的数据的数值分布、数据的字节位置改变概率、字节间的连锁改变概率以及字节位置的数据值分布进行了分析,并给出了定量化的实验结果。·针对数据值局部性比较弱的情况,本文提出了一种快速而高效的差分式写减少方法FEBRE。该方法采用了流水化伪随机编码算法和星形生成规则实现了一个一对多映射。实验表明该方法在同类算法中具有最好的写减少效果,以及较低的读写延迟和实现代价。·本文建立了差分式写减少方法在均匀分布数据集上的写入量模型,给出了不同字长-标志位组合下的最小写入量公式和计算结果。然后,本文提出了称为R2D2和R2D2-L的差分式写减少方法。前者具有接近理论极限的写减少能力。后者以较小的写减少能力损失为代价,取得了极低的实现代价。·针对数据值局部性较好的应用,本文提出了一种PCM寿命延长方法RMB。通过将某些存储芯片上过多的写入量转移到附加的存储芯片中,该方法避免了写平均方法对轻写入负载位置不够友好的缺点,且同时具备写平均和写减少的效果。实验表明,在某些情况下RMB方法甚至可以取得优于理想写平均方法的寿命延长效果。更多还原

【Abstract】 The current computing system has to employ memory hierachy,which combines several kinds of storage devices together,to fullfill the requirement of speed,thoughput,capacity,and non-volatility.However,this design brings more complexity to the hardware/software design and more power consumption.While the emerging phase change memory(PCM)technique,which is treated as a candidate for future memory system,is hoped to change the current situation because it has almost every property that a perfect storage device should have.However,PCM is suffered from the limited endurance problem.Therefore,the major purpose of the present work is trying to provide solutions for this problem.Specifically,the endurance of PCM is not only sensitive to the repeat pattern of writing address,but also to the pattern of writing content.So,the present work is started at the research on the pattern of content at first,which followed by two studies on reducing number of flipped bits per write and a research on how to balance write amount difference among storage locations.The achievements and contributions of the present work are as follows:· The present work discusses the data value locality in detail and points out that the locality is consisted of two behaviours: the macro behaviour and the micro behaviour.Based on experiments,the selected applications are run on X86-32,X86-64 and arm-32 environments.The information of their memory trace is collected.The data value,the change rate of byte location,the chained change rate of the byte location,and the value distribution on different byte localiton are analyzed quantitatively,as the basis of the following work.· The present work proposes a fast and efficient differential write reduction method,named FEBRE,serving for the applications with near-uniformly distributed writing data that is a sign of weak data value locality.The key points of the method include a pipelined pseudo-random encoding algorithm,an star-generating rule and a general encoding direction pattern to accelerate read/write operations and eliminate redundant elements in the encoded candidate set which could reduce more bit flipping.The experiment proves that FEBRE has the best write reduction capability among all known similar methods,and meanwhile,it keeps low read/write delay as well as the implementation overhead.· The present work furtherly discusses the theoretical limitation of the differential write method and the way to reach the limitation.By establishing a math model,the numeric values of the limitation,which means the minimum number of flipped bits per write,are calculated.Then,a novel differential write reduction method,named R2D2 is proposed,as well as its simplified implementation R2D2-L.The former,based on linear congruential pseudo-random function,is very close to the theoretical limitation of wrtie reduction.While the latter,R2D2-L,which is close to R2D2 on the write reduction capability,has a very low implementation overhead.· The present work proposes an endurance enhancement method,named RMB,which is trying to serve applications with strong data value locality.An interesting feature of the method is that the write reduction effect and wear leveling capability are both achieved.Specifically,Through redirecting the writing amount from the current heavy-wearing storage chip to the auxiliary long-life chip,RMB method achieves the both effects and meanwhile successfully avoids the disadvantage of wear leveling method,which is not friendly to the light wearing storage locations.The experiment shows that RMB method is able to perform better even than the perfect leveling method under some cases,which is the theoretical limitation of all wear leveling methods.更多还原