希捷瓦记录磁盘评测

分析了可提升磁盘存储密度的瓦记录(SMR)技术的优势及局限性,指出SMR采用部分叠加相邻磁道的方式增大磁盘的面密度,但是这也带来了写覆盖(Write Overlay)问题,即更新某一磁道上的数据时会覆盖相邻磁道上的有效数据,导致瓦记录磁盘(SWD)无法支持原地更新,使其非顺序写性能受到影响。为了促进该技术的应用,通过Fio和Filebench测试工具,对SMR技术的代表性产品——希捷瓦记录磁盘(SSWD)做了比较全面的测试,并对多种应用场景下SSWD的性能做了详细的测试和分析,以便为SSWD在数据中心的应用提供重要的参考。     

一种基于捎带回收的瓦记录Raid5写顺序化方法

瓦记录技术(SMR)采用部分叠加相邻磁道的方式增大磁盘碟片的面密度,其磁道相互叠加,使得瓦记录磁盘无法进行原地更新,是因为直接更新某一磁道上的数据可能会覆盖掉相邻磁道上的有效数据,导致瓦记录磁盘的非顺序写需要引入额外的写放大开销。对于由瓦记录磁盘组成的Raid5来说,在非顺序写场景下,最终落盘的IO也是非顺序的,因此单盘性能下降,会极大地影响瓦记录Raid5的非顺序写性能。针对瓦记录Raid5非顺序写性能低的问题,提出了一种基于捎带回收的瓦记录Raid5。该方法改进了传统Raid5的小写过程,在小写读的时候捎带读取附近的可被回收的数据或校验,通过将落盘写IO日志化,实现了落盘写IO的顺序化,避免了瓦记录磁盘的写放大开销。系统维护了地址映射表等元数据,在重构时只需对应用访问过的条带进行读写和计算,这减少了重构时需要读写的数据量。测试显示,在某些非顺序写场景下,写带宽超过了以Bcache作为缓存的MD Raid5。从计算结果看,重构时的读写开销相比普通磁盘组建的MD Raid5也有明显减小。     

基于区域协作的Cache压缩

为提高Cache的有效容量,进行了Cache压缩研究,并提出了一种区域协作压缩(RCC)方法,以提升最后一级缓存的压缩率。与传统的Cache压缩算法不同,RCC方法利用了缓存区域的压缩局部性,使用缓存区域中第一个缓存块的字典信息来协作压缩缓存区域中的其他各个缓存块,而不需要对缓存区域进行整体压缩。RCC有效发掘了缓存区域内缓存块之间的数据冗余,实现了接近以缓存区域为压缩粒度的字典压缩的压缩率,然而压缩、解压缩延时却仍然和压缩单个缓存块时相当。实验结果表明,与单缓存块压缩算法C-PACK相比,RCC方法的压缩率平均提升了12.34%,系统的性能提升了5%。与2倍容量的非压缩Cache相比,有效容量提升了27%,系统性能提升了8.6%,而面积却减少了63.1%。     

TH-MSNS系统中基于IP协议的缓存机制设计与实现

为提高IP-SAN的性能,在清华大学海量存储网络系统（TH-MSNS）存储区域网络的基础上,设计和实现了一种iSCSI环境下的缓存系统。该系统采用服务器内存作为数据的缓存,直接在内存中完成部分读写数据的命令,在存储请求空闲时同步远程网络磁盘数据。通过性能对比测试表明,该缓存系统能够较大地提高IP-SAN存储系统的性能,能增大存储系统的带宽,减少操作延迟。     

基于集群系统的空间数据并行处理策略研究

为了解决单节点的WebGIS系统存储能力和计算能力受限的问题,提出了一种利用分布式文件系统(DFS)和MapReduce分布式计算框架在集群环境中并行处理空间数据的方法.还特别针对分布式文件系统,结合WebGIS的应用模式,提出了小文件优化策略.该策略的核心思想是通过将小文件合并为大文件来有效降低文件的数目.试验结果表明,在使用了小文件优化策略后,分布式文件系统中的节点平均内存占用率从55.78%降至18.36%,文件的存储和读取性能分别提高了63.3倍和2.0倍.其次,基于经过优化后的分布式文件系统和MapReduce计算框架,设计并且实现了HDWebGIS原型系统,试验结果表明,使用了小文件优化策略后,HDWebGIS系统性能比优化前提升了78.11%.     

基于NVDLA与FPGA结合的神经网络加速器平台设计

随着深度神经网络对算力的需求不断增加,传统通用处理器在完成推理运算过程中出现了性能低、功耗高的缺点,因此通过专用硬件对深度神经网络进行加速逐步成为了深度神经网络的重要发展趋势。现场可编程门阵列(FPGA)具有重构性强、开发周期短以及性能优越等优点,适合用作深度神经网络的硬件加速平台。英伟达深度学习加速器(NVDLA)是英伟达开源的神经网络硬件加速器,其凭借自身出色的性能被学术界和工业界高度认可。本文主要研究NVDLA在FPGA平台上的优化映射问题,通过多种优化方案高效利用FPGA内部的硬件资源,同时提高其运行性能。基于搭建的NVDLA加速器平台,本文实现了对RESNET-50神经网络的硬件加速,完成了在ImageNet数据集上的图像分类任务。研究结果表明,优化后的NVDLA能显著提高硬件资源使用效率,处理性能最高可达30.8 fps,实现了较边缘中央处理器(CPU)加速器平台28倍的性能提升。     

面向稀疏卷积神经网络的CGRA加速器研究

本文针对规模日益增长和演变迅速的稀疏卷积神经网络(CNN)应用,提出一款高能效且灵活的加速结构DyCNN来提升其性能和能效。DyCNN基于兼具灵活性和高能效的粗粒度可重构架构(CGRA)设计,可以利用其指令的高并行性来高效支持CNN的操作。DyCNN使用基于数据感知的指令动态过滤机制来滤除各计算单元中由于稀疏CNN中权值静态稀疏性和激活值动态稀疏性产生的大量无效计算和访存指令,使它们能像执行稠密网络一样高效复用一组指令。此外DyCNN利用基于负载感知的动静结合负载调度策略解决了稀疏导致的负载不均衡问题。实验结果表明,DyCNN运行稀疏CNN与运行密集CNN相比实现了平均1.69倍性能提升和3.04倍能效提升,比先进的GPU(cuSPARSE)和Cambricon-X上的解决方案分别实现了2.78倍、1.48倍性能提升和35.62倍、1.17倍能效提升。     

基于访问预测的P2P VoD系统服务端数据缓存策略

分析了P2P VoD系统中各服务端数据访问顺序性与持续性的特点,提出了一种适用于P2P VoD系统的基于访问预测的服务端数据缓存策略(CAF).该策略对顺序访问与随机访问进行预测并实施数据预取,在计算P2P VoD网络中节点数据访问持续性的基础上,由历史访问与预测的未来访问推断缓存项未来被访问频率,以该频率选择缓存替换...     

UGD:面向固态盘缓存系统的数据过滤技术

为了提高固态盘(SSD)缓存系统固态盘的使用寿命,研究了系统的数据过滤技术.针对现有固态盘缓存系统数据过滤技术没有考虑数据页重用距离,导致重用距离大于固态盘缓存大小的数据页进入固态盘缓存,造成不必要的缓存页替换和固态盘擦写,从而降低缓存命中率和固态盘使用寿命的问题,提出了一种协同的数据过滤技术UGD.该技术综合考虑了数据页访问频率和数据页重用距离,让访问次数少和重用距离长的数据页不进入固态盘缓存,从而增长了固态盘的使用寿命.为了评测UGD技术的效果,实现了一个固态盘缓存模拟器,并使用开源负载和实际应用负载进行了大量实验.实验数据表明,与现有过滤技术相比,UGD技术能够使固态盘缓存系统的平均命中率提高10％,使固态盘平均写入量降低42.5％.     

多核处理器系统I/O访存优化研究

本文提出了一种多核处理器自适应I/O直接缓存访问(ADCA)的方法以提升I/O访存的性能,降低对其他程序的影响。与传统直接缓存访问(DCA)不同的是,该方法利用了LRU栈特性,通过采样辅助标签目录的方式动态调整DCA可使用的cache空间,同时对I/O数据的替换和写内存策略进行优化。实验结果表明,与DCA方式相比,该方式使得I/O带宽提升了大约10%,而与SPEC和采用直接内存访问(DMA)方式的网络测试程序同时运行相比,SPEC定点和浮点性能分别提升了11. 5%和8. 9%。     

L1₀ FePtCu bit patterned media

Chemically ordered 5 nm-thick L1? FePtCu films with strong perpendicular magnetic anisotropy were post-patterned by nanoimprint lithography into a dot array over a 3 mm-wide circumferential band on a 3 inch Si wafer. The dots with a diameter of 30 nm and a center-to-center pitch of 60 nm appear as single domain and reveal an enhanced switching field as compared to the continuous film. We demonstrate successful recording on a single track using shingled writing with a conventional hard disk drive write/read head.     

Cache Memory Design for Single Bit Architecture with Different SenseAmplifiers

Most modern microprocessors have one or two levels of on-chip caches to make things run faster, but this is not always the case. Most of the time, these caches are made of static random access memory cells. They take up a lot of space on the chip and use a lot of electricity. A lot of the time, low power is more important than several aspects. This is true for phones and tablets. Cache memory design for single bit architecture consists of six transistors static random access memory cell, a circuit of write driver, and sense amplifiers (such as voltage differential sense amplifier, current differential sense amplifier, charge transfer differential sense amplifier, voltage latch sense amplifier, and current latch sense amplifier, all of which are compared on different resistance values in terms of a number of transistors, delay in sensing and consumption of power. The conclusion arises that single bit six transistor static random access memory cell voltage differential sense amplifier architecture consumes 11.34 μW of power which shows that power is reduced up to 83%, 77.75% reduction in the case of the current differential sense amplifier, 39.62% in case of charge transfer differential sense amplifier and 50% in case of voltage latch sense amplifier when compared to existing latch sense amplifier architecture. Furthermore, power reduction techniques are applied over different blocks of cache memory architecture to optimize energy. The single-bit six transistors static random access memory cell with forced tack technique and voltage differential sense amplifier with dual sleep technique consumes 8.078 μW of power, i.e., reduce 28% more power that makes single bit six transistor static random access memory cell with forced tack technique and voltage differential sense amplifier with dual sleep technique more energy efficient.     

Design of a manufacturable discrete track recording medium

The potential benefits of patterning discrete tracks onto a disk for magnetic data storage have long been investigated. A practical process for manufacturing a cost-effective discrete track recording (DTR) medium has prevented such a disk from being introduced into a product. In this paper, a process utilizing nano-imprint lithography techniques to create a land and groove structure on the surface of a disk substrate will be described. Design considerations for the geometry of the structure, as well as of the magnetic write and read widths of the head, are discussed. Data showing the magnetic characteristics and recording performance of a DTR medium are also presented.     

Review Modern magnetic materials in data storage

The current status of the technology of magnetic recording as used in disk drives is reviewed. The emphasis is on the magnetic materials used in the application and on some of the technical problems that may limit the increase in areal density. The new technology of magnetic random access memory (MRAM), which has evolved from the magnetic recording application, is also reviewed. A wide range of magnetic materials is essential for the advance of magnetic recording and the MRAM technology. For the magnetic-recording application the requirements are for high-magnetization, soft magnetic materials for write heads, new antiferromagnetic alloys with high blocking temperatures, large coupling to ferromagnetic films and low susceptibility to corrosion for pinning films in giant magnetoresistive sensors, and for the MRAM application, the requirement is for new ferromagnetic alloys with large values of tunneling polarization ratio. A significant limitation to magnetic recording is found to be the inconsistent demands on media thickness: small media thicknesses are required for large values of signal-to-noise ratio, while large values of thickness are required to reduce the impact of the superparamagnetic effect, which results in the potential for data loss over time. Both of these requirements are discussed. Multilayer ferromagnetic films for recording surfaces are shown to allow both large signal-to-noise ratio and adequate resistance to data loss.     

采用图像格式的IDE硬盘直写记录

介绍一种采用IDE硬盘的高速图像记录系统的设计方法。该方法去掉了大量的软件编程代 之以硬盘直写图像文件的模式,将图像数据记录到硬盘的同时直接以文件形式存放。实验效果表明,采用该设计方法的记录设备做到了记录完全脱机,事后将硬盘通过USB接口直接与计算机相连,即可对图像进行访问,达到了预期的目的。     

Micromagnetics of Ferromagnetic Nano-Devices Using the Fast Fourier Transform Method

Micromagnetic simulations are widely utilized to simulate magnetic properties of magnetic materials, performance of magnetic devices, and read/write processes in recording systems. In hard disk drives, the minimum size of devices is now scaled down to the nanometer region; therefore, micromagnetics should be accurate enough to include the edge effects in these ferromagnetic nano-devices. In this work, we consider devices with flat structures. The analytical forms of demagnetizing matrices for intra-cell and inter-cell interactions are calculated for cubic and right-angle triangular prism cells respectively, which enables accurate magnetostatic interaction calculation using fast Fourier transform (FFT) method in thin film devices with arbitrary geometry. Besides, the parallel computational method is included to speed up the programs. Two examples about the switching characteristics of ferromagnetic devices are given in this paper to evaluate the accuracy of the improved FFT methods: one is a hard magnetic nano-dot; the other is a write pole tip for perpendicular recording.     

Recording performance of discrete track patterned media fabricated by focused ion beam

We report on the recording performance of discrete track patterned media fabricated by focused ion beam (FIB). We investigated performance over a small area by spinstand read/write testing. Discrete track patterned regions show smaller magnetic track width and better signal separation between adjacent tracks and therefore higher track density than that of nonpatterned continuous media as a result of reduced side fringe effect and edge noise. We found that, at a designed groove depth of 4-8 nm, the shallow FIB etched grooves already provide good isolation between adjacent tracks, indicating the superiority of ion beam induced modification of magnetic properties in film media over physical modification of disk surface topography. This has implications for discrete track recording and media fabrication.     

Performance study and analysis of dual-element head on thin-filmdisk for gigabit-density recording

Inductive-write and magnetoresistive-(MR)-read dual-element heads with very narrow tracks and gaps have been designed, fabricated, and tested on thin-film media of high coercivity and squareness. The results not only show excellent writeability at modest write currents but also the existence of a narrow region of optimum write current, limited by the onset of self-erasure by the write head at high write currents. This leads to significant degradations of overwrite, signal amplitude, trackwidth, linear resolution and disk-noise-induced peak-jitters. A peak-jitter approach is shown to be useful in characterizing many aspects of recording performance. A peak-jitter evaluation of signal-to-noise behavior reveals not only satisfactory overall performance but also the dominance of disk noise as well as a concentration of the disk noise at the track edges. Peak-jitter evaluations of offtrack and squeeze behavior clearly demonstrate the narrow-track capabilities of these recording heads for high areal density operation