分布式环境中的性能预测方法

发布时间：2018-08-09 15:22

【摘要】：在过去的十几年里，，分布式计算技术得到了广泛的研究和应用。在分布式系统中，用户共享所有的资源，彼此之间存在着竞争关系，为了提高分布式系统的性能，有效的资源分配机制显得格外重要。而准确的对资源使用情况进行预测可以使资源分配更加有效，所以本文将主要研究如何更准确的对各种资源使用情况进行预测。通常将系统资源使用情况看做时间序列进行分析和预测，传统的时间序列分析方法如自回归模型等都可以用于系统资源使用情况的分析和预测，近些年来更多的非线性模型被应用于时间序列的预测也取得了很好的效果。根据参考变量维度的不同，时间序列的预测可以分为单变量预测和多变量（多维度）预测。现有的各种预测模型都存在对数据的敏感性，往往在不同数据集上预测效果相差较大。另外实际情况下，在多变的分布式系统中，也很难保证特定机器上某种资源的变化规律一成不变。因而在本文中，我们提出了一种两层反馈式集成预测模型，一方面根据集成学习的思想提高预测的准确度与适应度，另一方面，不断对各个基础预测器进行优化，更进一步的提高预测能力。预测器优化模块使预测器集成模块获得更好的结果，同时预测器集成模块会根据集成的结果反作用于预测器优化模块，这种相互作用不断提高集成预测模型的预测能力。首先我们将该集成预测模型应用于单变量预测，将常用的几种单变量预测模型进行集成，并通过设计一系列的实验验证了集成预测模型的预测能力。接下来，我们介绍了几种多变量预测模型，在多变量预测中机器学习的方法取得了很好的效果，我们对其中的支持向量机回归预测模型进行了优化，然后与其他几种多变量预测模型一起构成我们的多变量集成预测模型。通过一系列的实验表明，该集成预测模型在多变量预测中同样有较为理想的效果。
[Abstract]:In the past decade, distributed computing technology has been widely studied and applied. In distributed systems, users share all resources, and there is a competitive relationship between them. In order to improve the performance of distributed systems, effective resource allocation mechanism is particularly important. And accurate prediction of resource use can make resource allocation more effective, so this paper will mainly study how to predict the use of various resources more accurately. The use of system resources is usually regarded as time series analysis and prediction. Traditional time series analysis methods such as autoregressive model can be used to analyze and predict the system resource use. In recent years, more nonlinear models have been applied to the prediction of time series. According to the different dimensions of reference variables, the prediction of time series can be divided into single variable prediction and multivariate (multivariate) prediction. All kinds of existing prediction models are sensitive to data, and the prediction results vary greatly in different data sets. In addition, in the changeable distributed system, it is difficult to ensure that the rule of change of a certain resource on a particular machine remains unchanged. Therefore, in this paper, we propose a two-layer feedback integrated prediction model. On the one hand, we improve the accuracy and fitness of prediction according to the idea of integrated learning; on the other hand, we constantly optimize each basic predictor. Further improve the ability to predict. The predictor optimization module makes the predictor integration module obtain better results, and the predictor integration module will react to the predictor optimization module according to the integrated results. This interaction improves the prediction ability of the integrated prediction model. Firstly, we apply the integrated prediction model to single variable prediction, and integrate several commonly used single variable prediction models, and design a series of experiments to verify the prediction ability of the integrated prediction model. Then, we introduce several kinds of multivariate prediction models. The machine learning method in multivariate prediction has achieved good results. We have optimized the support vector machine regression prediction model. Then our integrated multivariable prediction model is constructed with several other multivariable prediction models. A series of experiments show that the integrated prediction model is also effective in multivariate prediction.
【学位授予单位】：上海交通大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP338.8

【引证文献】