【作者】 刁莹；

【导师】 姚念民；

【作者基本信息】 哈尔滨工程大学 ， 计算机应用技术， 2013， 博士

【摘要】 作为信息化时代的重要资源，存储系统具有重大的实用价值。随着数字信息的爆炸性增长，大数量的用户群和种类繁多的应用导致了人们对大容量信息存储系统的需求。为了方便许多资源的使用，发展存储系统是首选的普遍的途径，所以存储系统是应用广泛的，并且在迅速地发展。在当今存储系统中，由于I/O性能是存储系统性能不可或缺的一部分，因此如何构造一个高性能、低能耗、适合不同应用的I/O系统就成为了一个重要的任务。存储系统的性能无论是对存储系统自身还是基于存储系统的应用都是至关重要的。其中I/O性能是评价存储系统的重要指标，对存储系统I/O性能的研究具有很大的理论和实用价值。目前计算机的I/O性能已经成为严重制约系统整体性能的瓶颈，所以I/O性能的分析与优化方法对于提升I/O性能就显得尤为重要。近年来，排队论方法已经成为分析和探讨存储系统性能的主要方法，因此本论文以排队论为主要工具来分析和研究有关存储系统I/O方面的性能。如今，在大规模存储系统中，由于数据的增长和大量存储设备的存在，使得几乎每天都会有存储设备发生故障，因此，存储设备的可修性也成为一个必须考虑的因素。影响存储系统性能的因素很多，本论文分析和研究的重点是前端排队方式对整个存储系统的影响。已有很多专家分析和研究了存储系统的总体特征，因此，本论文主要分析一些局部特征，而且是针对存储系统I/O的问题建立相应的模型来分析I/O的服务设备、结构布置、运作模式等等。本文使用排队论和拟生灭过程相结合的方法分析了I/O性能，主要是通过排队论的一些指数分布和拟生灭过程的Q矩阵方法，例如假设服务效率等参数服从负指数分布，提出了矩阵模式的分析方法，总结了I/O运行的具体相关性能指标表达式。然后进行仿真，经过多次仿真得到参数之间的规律关系，将用来仿真的数据用于I/O具体的操作过程中，最后总结I/O不同情况的操作规律以及所有运作模式的普遍规律，所得结果为I/O性能评价以及调度策略的确定提供了的依据。缓存是影响存储系统性能的重要因素。本文提出一种符合局部性原理的访问序列生成过程。这个过程具有局部性、延续性和突变性等良好性质，并且可以通过几个参数的设置控制以上各性质的强弱，生成符合各种访问模式的数字流，并对比各种缓存替换算法的优劣。在此方法的基础上，针对应用服务器和存储服务器构成的一个二级buffercache结构进行了研究。现有的ULC（Unified Level-Aware Caching）机制可有效解决多级缓存的数据冗余及在存储服务器端缓存的访问弱局部性问题。但是当存储服务器连接多个应用服务器时，ULC采用LRU策略为各应用服务器在存储服务器端分配缓存容量，该方法不能使存储服务器端缓存资源的边际收益最大化，为此本文提出一种多应用共享缓存的二级buffer cache动态分配策略MG-ULC（Marginal Gain-based ULC）。该策略以ULC机制为基础，根据各应用的访问模式在二级buffer cache的边际增益动态分配缓存容量。实验结果显示，随着各应用服务器访问模式的变化，MG-ULC能比ULC更合理地分配二级缓存，从而实现更高的缓存利用率，使存储系统性能进一步提升。I/O负载是影响存储系统性能的重要因素，响应时间是反映存储系统性能的重要指标。为了提高存储系统的性能，本文对I/O负载的特征参数和响应时间进行了研究分析。为实现根据I/O负载的一些特征参数预测系统响应时间的目的，本文利用灰色系统理论和BP人工神经网络相结合的方法建立模型，确定了特征参数和响应时间之间的非线性关系，利用disksim得出trace数据，进行仿真，并将预测结果和单纯使用灰色和神经网路预测的结果进行比较，突出其优越性。由于考虑的因素不同，本文又结合BP神经网络和马尔科夫链，建立了一种新的用于I/O负载的预测模型BP神经网络-马尔科夫链预测模型（BP-MC）。通过对训练样本的学习，利用BP神经网络实现了对负载时间序列的滚动预测，同时得到了实测值与预测值的相对误差。在此基础上利用马尔科夫链对相对误差进行修正，有效的提高了预测结果的精度。将该模型应用于存储系统I/O负载预测中，结果表明该模型预测精度高，为存储系统性能预测提供了新的途径。I/O性能评价在存储系统的设计和应用中占有非常重要的位置，如何快速有效地评价I/O的性能是优化I/O性能的一个重要步骤。本文提出了用定性和定量相结合的方法，即层次分析方法和人工神经网络方法相结合的方法来分析和研究I/O的性能，同样用disksim得出的trace数据进行仿真，仿真结果显示了其优越性，通过实例说明了此评价方法可以有效地评价存储系统的I/O性能。更多还原

【Abstract】 As an important resource, the storage system is of great significance. With the explosivegrowth of digital information, users and various applications, the demands of high-capacitystorage systems is increasing very fast. The development of the storage system is thepreferred and prevalent way for the good use of many resources.In today’s storage system, the performance of I/O system is an essential part of theperformance of storage system. Therefore, how to construct an I/O system with highperformance, low power consumption and suitable for different applications becomes animportant task. The performance of storage system is vital to the storage system itself and theapplication of storage systems. I/O performance is one of the most important performanceindicators for evaluating the storage system, and it has great theoretical and practical value forthe study of I/O performance of storage system. At present, because I/O performance ofcomputer has already become the bottleneck of the whole system for a long time, I/Operformance analysis and optimization method looks especially important and valuable.In recent years, as the main methods of storage system performance, the queuing theorymethod has been analyzed and discussed. Therefore, in the thesis, the queuing theory is takenas the main tool to analyzes and study the performance of I/O aspects of storage system.Because of the growth of information and the existence of many storage devices in thelarge-scale storage nowadays, the storage device failures can occur almost every day, thus, therepairability of storage devices must to be a considerable factor. There are too many factorsinfluencing storage system performance. The paper focuses on the front-end queue of thewhole storage system.This thesis mainly analyzes some local characteristics, and aiming atthe I/O problem of storage system, the service equipment, structural arrangement, andoperation model of the I/O and so on are analyzed by establishing corresponding model. Inthe thesis, I/O performance model is analyzed by the use of the method of the combination ofquasi birth and death process and queuing model, and it is studied mainly passing theindicator distribution of queuing theory and the Q matrix pattern of quasi birth and deathprocess, such as service efficiency submit to megative exponential distribution and so on. Theanalytical method of matrix pattern is presented, and the I/O operational specific performanceindicator expression is summed up. Then after many simulations, the regular pattern isobtained. The data are use for the I/O concrete operation process, finally, the I/O operation rule under the different conditions and all the universal law under the operation mode aresummarized. The results may be used to evaluate storage performance, and are the basis fordeciding I/O scheduling strategy.Cache is one of the important factors affect the storage system performance. This paperproposes a visit sequence generation process comply with the principle of locality. Theprocess has locality, continuity and mutability properties and so on, and can control thestrength of the above properties by a few parameters set. The digital flow is generated in linewith the various access patterns, and compared with various cache replacement algorithm. Onthe basis of this method, a two-level buffer cache structure that is constituted by the buffercaches of application servers and storage servers is studied. The existing ULC (UnifiedLevel-Aware Caching) protocol can effectively solve the problems that redundantly cachedblocks in multilevel hierarchy and weakly localized at storage server cache. However, whenthere are multiple application servers sharing one storage server, the ULC adopts LRUstrategy to allocate cache capacity of storage server to each application server, and thismethod can not gain the maximal marginal profits of the storage server cache. A second-levelbuffer cache dynamic allocation strategy called MG-ULC (Marginal Gain-based ULC) isproposed, and it is designed for storage servers in which multiple applications share the samecache resources. Based on the ULC protocol, the MG-ULC dynamically allocates cachecapacity in accordance with the second-level buffer cache marginal gain of each application.The results shows that, as each application’s access pattern changes, the MG-ULC canallocate second-level buffer cache more rationally than the ULC, thereby realizing a highercache utilization.I/O load is an important factor to affect the performance of storage system, and theresponse time is an important indicator to reflect the performance of storage system. In orderto improve the performance of storage system, the characteristic parameters of the I/O loadand response time are analyzed in this paper. In order to achieve the purpose of forecastingresponse time by some characteristic parameter of I/O load, the model is built by combiningwith the Grey System Theory and BP Artificial Neural Network method, and the nonlinearrelations of characteristic parameter and response time is determined. Make use of the disksim,the trace data is obtained, then the trace data is used in simulating. The predicted result iscompared with the results of using grey and neural network prediction, and the advantages arehighlighted. Due to the considered factors are different, a new prediction model is proposedby combining with the BP neural network and Markov Chain, named as BP neural network-Markov chain(BP-MC) model and applied to I/O load prediction. Through emulatingthe training sample, the rolling prediction of load time series is achieved by BP neuralnetwork, and the relative error of measured and predicted values is acquired. Applying theMarkov Chain to correct the relative error and the accruacy of predicting results is improvedeffectively. The prediction model was used to predict the I/O load of Storage system， theresult shows that the model is high-precision， so it provides a new approach for Storagesystem prediction.The evaluation of I/O performance is very important in the design and use of storagesystem. How to evaluate the I/O performance quickly and efficiently has become an importantprocedure for optimizing I/O performance. In this thesis, both qualitative and quantitativemethod is proposed, which is the combination of analytical hierarchy process level analysismethod and artificial neural network method. I/O performance is analyzed and studied by thismethod. Similarly, the trace data used for simulation is obtained by making use of the disksim.The advantage is found from the simulate result. The evaluation methodology can effectivelyevaluate the I/O performance of storage system.更多还原