基于日志分析的超级计算机错误预测方法研究*

大量的错误严重影响了超级计算机系统的稳定性,错误预测对于提高其稳定性有重要作用,日志分析是进行错误预测的有效方法.建立了错误预测的基本框架,包括日志的预处理、基础预测器和联合预测器,其中基础预测器包括时间预测器和关联预测器.在BlueGene/L日志上进行的实验结果显示联合预测器的预测效果比基础预测器好.这表明错误预测...     

基于Markov模型与轨迹相似度的移动对象位置预测算法

针对低阶Markov模型预测精度较差,以及多阶Markov模型预测稀疏率高的问题,提出一种基于Markov模型与轨迹相似度(MMTS)的移动对象位置预测算法。该方法借鉴了Markov模型思想对移动对象的历史轨迹进行建模,并将轨迹相似度作为位置预测的重要因素,以Markov预测模型的预测结果集作为预测候选集,结合相似度因素得出最终预测结果。实验结果表明,与k阶Markov模型相比,该方法的预测性能不会随着训练样本大小及阶数k的变化受到很大的影响,并且在大幅降低k阶Markov模型预测稀疏率的同时将预测精度平均提高了8%以上。所提方法不仅解决了k阶Markov模型的预测稀疏率高及预测精度不足的问题;同时提高了预测的稳定性。     

用于图像预测编码的预测器的设计及应用

设计并实现了一个用于图像预测编码的预测器。在图像压缩中，预测器是影响整个图像压缩比以及图像质量的关键环节。为了提高预测器的性能，本文结合了线性与非线性预测器的设计方法，首先对图像像素及其邻域像素之间的相关性估计值进行统计计算．确定该预测器的最佳阶数为四阶，接下来以预测差值的最小均方差标准和差值图像的期望值恒为零的约束为前提，同时对邻域模板的平坦程度加以分类，统计分析各类图像得出大量实验数据，最终设计出性能较好的四阶预测器。由于该预测器应用于星载图像的压缩．其预测系数在设计时兼备了便于硬件实现的特点。     

基于SimpleScalar的动态分支预测器研究

分支预测精度是影响当代处理器性能的重要指标,在近十年内一直是学术界和工业界的研究热点。为给不同应用场合的处理器动态分支预测器设计提供性能参考,针对处理器架构设计中应用广泛的几种动态分支预测器,使用SPEC CPU2000在SimpleScalar模拟器中进行仿真及测试分析。测试结果以预测精度和指令/时钟周期作为指标,并结合硬件开销,分析了不同种类分支预测器的适用对象和场合。     

用于图像预测编码的预测器的设计及应用

设计并实现了一个用于图像预测编码的预测器。在图像压缩中,预测器是影响整个图像压缩比以及图像质量的关键环节。为了提高预测器的性能,本文结合了线性与非线性预测器的设计方法,首先对图像像素及其邻域像素之间的相关性估计值进行统计计算,确定该预测器的最佳阶数为四阶,接下来以预测差值的最小均方差标准和差值图像的期望值恒为零的约束为前提,同时对邻域模板的平坦程度加以分类,统计分析各类图像得出大量实验数据,最终设计出性能较好的四阶预测器。由于该预测器应用于星载图像的压缩,其预测系数在设计时兼备了便于硬件实现的特点。     

基于自适应Kalman预测器的运动估计算法

利用图像序列估计目标运动速度是机器人视觉中的一项重要研究内容。它应用在机器人操作、导航、视觉跟踪等多项领域中。这些应用一般均要求运动估计算法具有较好的实时性和抗噪能力。卡尔曼滤波器和预测器正符合上述要求。该文基于运动图像的仿射模型，探讨从序列图像中预测目标三维平动速度的卡尔曼预测算法。首先建立运动目标的“当前”统计模型，然后根据运动图像的仿射模型找出图像运动参数与目标三维速度间的关系(图像运动参数由目标图像的几何矩计算获得)。最后结合自适应卡尔曼滤波和卡尔曼一步预测算法设计自适应卡尔曼一步预测器。为减轻预测器的发散性，对初始状态进行估计。仿真结果表明，基于“当前”统计模型和运动图像仿射模型设计出的自适应卡尔曼一步预测器具有较高的精度。     

基于Venn-Abers预测器的系统日志异常检测方法

为了更好地解决系统日志异常检测问题，引入一种对预测结果进行可靠性评估的统计学习算法Venn-Abers预测器。与传统的基于静态阈值的系统日志异常检测模型仅输出正常或异常的预测结果不同，Venn-Abers预测器会对预测结果进行概率评估。根据逻辑回归(LR)、支持向量机(SVM)、随机森林(RF)这三种基础算法，分别开发三种Venn-Abers预测器，其中基于SVM的Venn-Abers预测器将召回率从81%提高到94%,同时对Venn-Abers预测器的概率值计算过程进行了改进，使其运行效率显著提高。实验结果表明，三种Venn-Abers预测器与其基础算法相比，可以通过动态调整阈值，取得更加准确的异常检测结果。     

Adaboost算法改进BP神经网络预测研究

针对传统BP神经网络容易陷入局部极小、预测精度低的问题,提出使用Adaboost算法和BP神经网络相结合的方法,提高网络预测精度和泛化能力。该方法首先对样本数据进行预处理并初始化测试数据分布权值;然后通过选取不同的隐含层节点数、节点传递函数、训练函数、网络学习函数构造出不同类型的BP弱预测器并对样本数据进行反复训练;最后使用Adaboost算法将得到的多个BP神经网络弱预测器组成新的强预测器。对UCI数据库中数据集进行仿真实验,结果表明本方法比传统BP网络预测平均误差绝对值减少近50%,提高了网络预测精度,为神经网络预测提供借鉴。     

基于类型预测的甚块预测器

高性能的甚块预测器是保证EDGE体系结构性能的关键手段.为研究性能更好的甚块预测器,文中通过仿真实验发现甚块的出口类型独立于甚块的出口个数和甚块的动态执行结果而存在.以此为据,提出了基于类型预测的甚块预测器.该预测器摈弃了甚块出口号,直接对甚块出口类型进行预测.随后,根据对甚块出口类型可预测性的分析,通过实验证明甚块出口类型与历史和路径信息相关.仿真结果显示,与经典的基于出口预测的甚块预测器相比,文中提出的基于类型预测的甚块预测器能够将每千条指令误预测次数平均降低约10％.     

Markov逻辑网在链接预测中的应用*

链接预测是对实体间的关系进行预测,是一个重要而复杂的任务。传统同类独立同概率分布的方法会带来很大的噪音,导致预测效果很差。将Markov逻辑网应用到链接预测中,旨在改善这一问题。Markov逻辑网是将Markov网与一阶逻辑结合的统计关系学习方法。利用Markov逻辑网构建关系模型,对实体之间是否存在链接关系以及当链接关系存在时预测此链接关系的类型。针对两个数据集的实验结果显示了采用Markov逻辑网模型要比传统链接预测模型有更好的效果,进而为Markov逻辑网解决实际问题提供了依据。     

Cost effective mixed-type value predictor using distributed classification method

We have investigated the conventional mixed-type value predictors, pointing out their limitations due to the inefficient use of data table entries. To improve the cost effectiveness of the conventional mixed-type value predictors in terms of the performance/cost ratio, we propose a new mixed-type value predictor, which uses the distributed classification method. The proposed value predictor has no centralized classification tables, but it uses distributed and local classification tables for each subsidiary predictor to classify instructions, to update data tables, and to predict result values. Static analysis of the cost reveals that the proposed value predictor decreases the cost by 30 and 10% compared with two conventional predictors, respectively. As well, the proposed value predictor increases the performance by 1% in terms of IPC, and, finally, improves the performance/cost ratio by 40 and 10% compared with two conventional methods.     

Microbenchmarks for determining branch predictor organization

In order to achieve an optimum performance of a given application on a given computer platform, a program developer or compiler must be aware of computer architecture parameters, including those related to branch predictors. Although dynamic branch predictors are designed with the aim of automatically adapting to changes in branch behavior during program execution, code optimizations based on the information about predictor structure can greatly increase overall program performance. Yet, exact predictor implementations are seldom made public, even though processor manuals provide valuable optimization tips. This paper presents an experimental flow with a series of microbenchmarks that determine the organization and size of a branch predictor using on‐chip performance monitoring registers. Such knowledge can be used either for manual code optimization or for design of new, more architecture‐aware compilers. Three examples illustrate how insight into exact branch predictor organization can be directly applied to code optimization. The proposed experimental flow is illustrated with microbenchmarks tuned for Intel Pentium III and Pentium 4 processors, although they can easily be adapted for other architectures. The described approach can also be used during processor design for performance evaluation of various branch predictor organizations and for testing and validation during implementation. Copyright © 2004 John Wiley & Sons, Ltd.     

基于神经网络的非线性Smith预估器

将Ｓｍｉｔｈ预估原理推广到非线性系统,并用神经网络予以实现对非线性磊滞后系统进行纯滞后补偿,并将所提出的神网络非线性Ｓｍｉｔｈ预估器用于非线笥动态系统即汽车发同歧气管压力系统的预报,并与基于神经网络的迭代及非迭代ｄ步超前预报器进行了比较,还给民基于机理模型的预报器的比较结果。     

A Novel Probabilistic Saturating Counter Design for Secure Branch Predictor

In a modern processor,branch prediction is crucial in effectively exploiting the instruction-level parallelism for high-performance execution.However,recently exposed vulnerabilities reveal the urgency to improve the security of branch predictors.The vital cause of the branch predictor vulnerabilities is that the update strategy of the saturating counter is deterministic.As a fundamental building block in a modern branch predictor,previous studies have paid too much attention to the performance and hardware cost and ignored the security of saturating counter.This leaves attackers with the opportunities to perform side-channel attacks on the branch predictor.This paper focuses on the saturating counter to explore a secure and lightweight design to mitigate branch predictor side-channel attacks.Instead of applying the isolation mechanism to branch predictor resources,we propose a novel probabilistic saturating counter design to confuse the attacker's perception of the victim's behaviour.It changes the conventional deterministic state transition function to a probabilistic state transition function.When a branch is committed,the conventional saturating counter needs to be updated about whether the prediction results are correct or not.While for the probabilistic saturating counter,the branch predictor determines whether the update is performed based on the update probability.The probabilistic saturating counter dramatically reduces the ability of the attacker to spy the saturating counter's state.Our analyses using a cycle-accurate simulator suggest that the proposed mechanism incurs 2.4％ performance overhead and hardware cost while providing strong protection.     

Stochastic theory of minimal realization

In this paper it is shown that a natural representation of a state space is given by the predictor space, the linear space spanned by the predictors when the system is driven by a Gaussian white noise input with unit covariance matrix. A minimal realization corresponds to a selection of a basis of this predictor space. Based on this interpretation, a unifying view of hitherto proposed algorithmically defined minimal realizations is developed. A natural minimal partial realization is also obtained with the aid of this interpretation.     

Virtual private network bandwidth management with traffic prediction

Dynamic link resizing is an attractive approach for resource management in virtual private networks (VPNs) serving modern real-time and multimedia traffic. In this paper, we assess the use of linear traffic predictors to dynamically resize the bandwidth of VPN links. We present the results of performance comparisons of three predictors: Gaussian, auto-regressive moving average (ARMA) and fractional auto-regressive integrated moving average (fARIMA). The comparisons are based on the mean packet delay, the variance of the packet delay, and the buffer requirements. Guided by our performance tests, we propose and evaluate a new predictor for link resizing: linear predictor with dynamic error compensation (L-PREDEC). Our performance tests show that L-PREDEC works better than Gaussian, ARMA and fARIMA in terms of the three metrics listed above. The benefit of L-PREDEC over the Gaussian predictor is demonstrated in two configurations: a common queue with aggregate link resizing and multiple queues with separate link resizing. In both configurations, L-PREDEC has consistently achieved better multiplexing gain and higher bandwidth utilization than its Gaussian counterpart.     

Finite dimensional predictors for hidden Markov chains

A continuous time Markov chain is observed in Gaussian noise. Finite dimensional normalized and unnormalized (Zakai) predictors are obtained for the state of the chain, for the number of jumps from one state to another and for the occupation time in any state.     

Observer‐predictor feedback for consensus of discrete‐time multiagent systems with both state and input delays

This article is concerned with the consensus problem for discrete‐time multiagent systems with both state and input delays. Single observer‐predictor‐based protocols and multiple observer‐predictors feedback protocols are simultaneously established to predict the future state such that the input delay that can be arbitrarily large yet bounded is completely compensated. It is shown that the consensus of the multiagent system can be achieved by the single/multiple observer‐predictors feedback protocol. Moreover, sufficient conditions guaranteeing the consensus of the multiagent system are provided in terms of the stability of some simple observer‐error systems, and the separation principle is discovered. Finally, a numerical example is worked out to illustrate the effectiveness of the proposed approaches.     

自校正对角阵加权信息融合Kalman预报器

For the multisensor systems with unknown noise statistics, using the modern time series analysis method, based on on-line identification of the moving average (MA) innovation models, and based on the solution of the matrix equations for correlation function, estimators of the noise variances are obtained, and under the linear minimum variance optimal information fusion criterion weighted by diagonal matrices, a self-tuning information fusion Kalman predictor is presented, which realizes the self-tuning decoupled fusion Kalman predictors for the state components. Based on the dynamic error system, a new convergence analysis method is presented for self-tuning fuser. A new concept of convergence in a realization is presented, which is weaker than the convergence with probability one. It is strictly proved that if the parameter estimation of the MA innovation models is consistent, then the self-tuning fusion Kalman predictor will converge to the optimal fusion Kalman predictor in a realization, or with probability one, so that it has asymptotic optimality. It can reduce the computational burden, and is suitable for real time applications. A simulation example for a target tracking system shows its effectiveness.