基于Kubernetes云原生的弹性伸缩研究

随着云技术的不断发展和普及,为了更好地利用云平台的优点和特性,云原生应用服务不断涌现,如何利用云平台的特性来服务软件设计和开发成为了难题,例如如何利用云平台的弹性伸缩特性。云原生目前主流的容器编排技术Kubernetes支持自动伸缩,却存在一些需要针对具体情况进行优化改进的问题。本文主要针对使用Kubernetes编排的5G核心网网元PCF(Policy Control Function)的水平自动伸缩进行研究,通过基于自定义的负载数据(CPU使用率、内存使用率、交易量、带宽使用率)统计,根据历史负载数据使用LSTM来预测未来的负载,并设计了一种基于预测负载的可行的弹性伸缩算法,从而提出一种提前感知的、弹性的、不影响业务的弹性伸缩方法,并进行了大量的实验和统计,来论证方法的可行性和正确性。     

基于多智能体强化学习的微服务弹性伸缩方法

针对现有微服务水平扩展策略难以应对异构应用对多种资源的差异化需求问题,提出了一种基于多智能体强化学习的微服务弹性伸缩方法。首先,通过刻画微服运行状态、资源调整动作及收益等要素建模云应用资源调整问题;其次,基于深度神经网络训练策略网络以决策资源调整操作,训练价值网络以评价决策优劣并优化调整策略;最后,提出中心化模型训练与分布式资源调整动作相结合的微服务弹性伸缩策略。实验结果表明,该方法能够根据负载波动及时调整各微服务的资源分配量,有效减少了云应用请求响应时间,并降低了云平台的资源使用成本。     

一种BP神经网络和遗传算法混合的云平台负载 预测方法

在云计算提供高效,便捷等强大服务的背后,是日益攀升的能耗问题。准确的预测云平台的负载(如CPU,内存的使用)在任务调度,云能效方面具有重要意义。在以往研究中,线性自回归算法在预测请求资源的粒度上存在不足,本文提出一种基于BP神经网络与遗传算法混合的负载预测方法,结合遗传算法良好的全局搜索能力与神经网络强大的非线性拟合能力,建立CPU资源的请求预测模型。实验通过Google的云平台数据作为训练,测试集。实验结果表明该方法有效的预测了CPU资源请求量,进而可以在此基础上调整服务资源,实现绿色调度。     

面向服务的云-端动态协作方法

边缘计算可以通过将计算转移至边缘设备,以提高大型物联网流数据的处理质量并降低网络运行成本。然而,实现大型流数据云计算和边缘计算的集成面临两个挑战。首先,边缘设备的计算能力和存储能力有限,不能支持大规模流数据的实时处理。其次,流数据的不可预测性导致边缘端的协作不断地发生变化。因此,有必要实现边缘服务和云服务之间的灵活划分。提出一种面向服务的云端与边缘端的无缝集成方法,用于实现大规模流数据云计算和边缘计算的协作。该方法将云服务分成两部分,分别在云端和边缘端上运行。同时,提出了一种基于改进的二分图动态服务调度机制。当产生事件时,可以在适当的时间将云服务部署到边缘节点。基于真实的电能质量监控数据对提出的方法进行了有效性验证。     

一种基于机器学习的虚拟机放置方法

为改善云数据中心的能耗、负载均衡性和服务等级协议(SLA)违背率,对虚拟机放置策略进行优化。基于IaaS环境,提出一种基于机器学习的虚拟机迁移调整方法。根据资源消耗的互补性和不均衡性对虚拟机进行预放置,使用深度神经网络预测物理机负载等级,并利用深度Q网络调整物理机数量。实验结果表明,该方法能够有效均衡负载分布,降低能源开销和SLA违背率。     

基于语义的服务混合匹配方法研究

为了提高Web服务匹配的效率和灵活性,提出了一种基于语义的Web服务混合匹配方法。该方法首先扩展了服务的功能描述模型,并基于该扩展模型进行混合服务匹配。混合匹配方法联合使用了两种匹配策略：基于推理的匹配策略,通过本体概念的逻辑关系实现匹配;基于相似性计算的策略,通过本体概念间相似度计算进行服务的匹配。在服务匹配过程中,首先使用基于推理的匹配策略将服务分成4个匹配等级,然后,采用基于相似性计算的匹配策略对候选服务进行进一步地匹配。实验结果表明该方法是可行而有效的。     

一种云环境下的主机负载预测方法

云计算是一种基于信息网络的计算模式和服务模式,它将信息技术资源以服务方式动态、弹性地提供给用户,使用户可以按需使用。由于受到主机的启动时间、资源分配时间以及任务调度时间等因素的影响,在云环境下提供给用户的服务存在时延问题。因此,工作负载预测是云环境下一种重要的能源优化的方式。此外,由于云中工作负载的变化具有十分大的波动性,因此增加了预测模型的预测难度。提出了一种基于自回归模型和Elman神经网络的预测模型(Hybrid Auto Regressive Moving Average model and Elman neural network,HARMA-E),其使用ARMA模型进行预测,再使用ENN模型对ARMA模型的误差进行预测,通过修正ARMA的输出值得到最终的预测值。仿真实验结果表明,该预测模型能够较好地提升主机负载预测值的准确度。     

机加工行业工业云边缘设备功能及应用

随着工业互联网的快速发展,机加工行业的设备和工业云平台之间的通信是一项重要的研究内容。针对不同的终端设备和云平台通信需要设置不同的通信流程这一问题,提出使用云边缘设备进行转接。边缘设备既能够和终端设备通信,又能够和云平台通信,同时还兼具计算和储存能力。该设备可以简化通信流程,提高通信效率。根据机加工行业的特点,设计了工业云边缘设备的部署框架,可以充分调用边缘设备的功能,将云平台的模型参数传递给边缘设备。这种方式扩展了边缘设备的事务处理能力,使得该边缘设备不仅能够处理简单的事务,还能处理复杂的事务,比如故障诊断、寿命预测等。最后,以计算数控机床健康度场景为例,验证了该边缘设备的功能以及部署框架的可行性。     

基于概率性能感知演化博弈策略的“云+边”混合环境中任务卸载方法

针对“中心云服务器+多个边缘服务器”构成的“云+边”混合环境中多任务卸载效率不足的问题,提出了一种基于概率性能感知演化博弈策略的任务卸载方法。首先,在一个“中心云服务器+多个边缘服务器”构成的“云+边”混合环境中,假设其中分布的边缘服务器具有时变波动的性能,采用一种基于概率性能感知演化博弈策略的任务卸载方法对边缘云服务器的历史性能数据进行概率分析,以获得演化博弈模型;然后,生成服务卸载的演化稳定策略（ESS）,使每个用户都能在获得高满意度的前提下进行任务的卸载。基于云边缘资源位置数据集和云服务性能测试数据集进行模拟实验,在24个连续时间窗口上进行不同方法的测试比较。实验结果表明,所提方法在多个性能指标上都优于传统的贪婪（Greedy）算法、遗传算法（GA）和基于纳什均衡的博弈论算法等任务卸载方法。该方法的平均用户期望达成度相较于三个对比方法分别提升了13.7%、117.0%、13.8%,平均卸载时延分别降低了6.5%、24.9%、8.3%,平均货币成本分别降低了67.9%、88.7%、18.0%。     

多源信息融合的微服务化电网事故追忆

提出一种在云计算平台上构建大数据环境下支持多源信息融合的微服务化电网事故追忆系统方法,解决传统电网事故追忆系统耦合度高、灵活性差、不易扩展等问题,满足事故追忆系统对短时间内解析大规模故障数据的要求。根据功能需求,将系统重构为电网模型、事故记录、操作记录和人机交互四个细粒度微服务;在此基础上,将每一个数据源接入独立的微服务模块,避免服务间的耦合和堆叠。结合云平台监控系统采集的容器集群负载参数,提出了基于长短期记忆神经网络算法的资源预测模型,提前对容器资源进行预测调度,避免负载突变对系统效率的影响,提升容器资源调度水平。结果分析表明,采用所述方法实现的微服务化电网事故追忆系统可靠性达99.999980%,具有良好的响应效率。     

基于组合预测模型的云计算资源负载预测研究

随着云计算技术的不断发展,云计算资源负载变化呈现出越来越复杂的特征。针对云计算资源的负载预测问题,综合考虑云计算环境中资源负载时间序列的线性与非线性特性,提出了一种基于自回归移动平均模型ARIMA与长短期记忆网络LSTM的组合预测模型LACL。使用公开数据集与传统负载预测模型进行了对比实验,实验结果表明,该云计算资源组合预测模型预测精度明显高于其他预测模型,显著 降低了云环境中对资源负载的实时预测误差。     

一种在线实时微服务调用链异常检测方法

微服务架构逐渐成为大规模云应用的主流设计架构，微服务可靠性是云服务商亟须处理的关键问题。精确检测并定位微服务应用故障可有效保障应用的可靠性与稳定性，基于微服务调用链的异常检测可在系统发生故障时及时发现系统异常行为并触发告警。针对当前主流检测方法无法保证异常告警的实时性和准确性问题，提出一种基于自然语言处理与双向长短期记忆（BiLSTM）网络的微服务调用链异常检测方法MicroTrace。对调用链中记录的事件进行解析，将事件表示为语义序列与响应时间序列，利用词汇嵌入式表示算法提取事件的向量化表示，通过基于注意力机制的BiLSTM同时检测微服务实例的调用路径与性能异常。在真实微服务调用链数据集上的实验结果表明，该方法的查准率和查全率均可达96%以上，F1度量值相比于多模态-LSTM方法至少提升了6.8%。     

Task offloading in cloud-edge collaboration-based cyber physical machine tool

The Cyber-Physical Machine Tool (CPMT) is a promising solution for the next generation of machine tool digitalization and servitization due to its excellent interconnection, intelligence, adaptability, and autonomy. The rapid development of next-generation information technologies, such as the Internet of Things (IoT) and artificial intelligence (AI), provided richer services for CPMT but also led to problems of idle on-site computing resources, and excessive pressure on the cloud, slow service response and poor privacy. To solve the above problems, this paper proposes a cloud-edge collaboration-based CPMT architecture, which makes full use of the computing resources of existing devices in the industrial sites, offloads digital twin (DT) modeling and data processing from the cloud to the edge, and provides microservice interfaces for users at the edge. Given the limited computing resources available in the field and the demand for latency-sensitive applications, task offloading methods aimed at response speed and load balancing are proposed, respectively. Finally, a case of machine tool Prognostics and Health Management (PHM) service is presented, in which the proposed method is used to perform tool wear monitoring, prediction, and health management.     

A self-learning fuzzy approach for proactive resource provisioning in cloud environment

The development of a communication infrastructure has made possible the expansion of the popular massively multiplayer online games. In these games, players all over the world can interact with one another in a virtual environment. The arrival rate of new players to the game environment causes fluctuations and players always expect services to be available and offer an acceptable service-level agreement (SLA), especially in terms of response time and cost. Cloud computing emerged in the recent years as a scalable alternative to respond to the dynamic changes of the workload. In massively multiplayer online games applications, players are allowed to lease resources from a cloud provider in an on-demand basis model. Proactive management of cloud resources in the face of workload fluctuations and dynamism upon the arrival of players are challenging issues. This paper presents a self-learning fuzzy approach for proactive resource provisioning in cloud environment, where key is to predict parameters of the probability distribution of the incoming players in each period. In addition, we propose a self-learning fuzzy autoscaling decision-maker algorithm to compute the proper number of resources to be allocated to each tier in the massively multiplayer online games by applying the predicted workload and user SLA. We evaluate the effectiveness of the proposed approach under real and synthetic workloads. The experimental results indicate that the proposed approach is able to allocate resources more efficiently than other approaches.     

Online machine learning for auto-scaling in the edge computing

The evolution of edge computing devices has enabled machine intelligence techniques to process data close to its producers (the sensors) and end-users. Although edge devices are usually resource-constrained, the distribution of processing services among several nodes enables a processing capacity similar to cloud environments. However, the edge computing environment is highly dynamic, impacting the availability of nodes in the distributed system. In addition, the processing workload for each node can change constantly. Thus, the scaling of processing services needs to be rapidly adjusted, avoiding bottlenecks or wasted resources while meeting the applications’ QoS requirements. This paper presents an auto-scaling subsystem for container-based processing services using online machine learning. The auto-scaling follows the MAPE-K control loop to dynamically adjust the number of containers in response to workload changes. We designed the approach for scenarios where the number of processing requests is unknown beforehand. We developed a hybrid auto-scaling mechanism that behaves reactively while a prediction online machine learning model is continuously trained. When the prediction model reaches a desirable performance, the auto-scaling acts proactively, using predictions to anticipate scaling actions. An experimental evaluation has demonstrated the feasibility of the architecture. Our solution achieved fewer service level agreement (SLA) violations and scaling operations to meet demand than purely reactive and no scaling approaches using an actual application workload. Also, our solution wasted fewer resources compared to the other techniques.     

E2LG: a multiscale ensemble of LSTM/GAN deep learning architecture for multistep-ahead cloud workload prediction

The Journal of Supercomputing - Efficient resource demand prediction and management are two main challenges for cloud service providers in order to control dynamic autoscaling and power consumption...     

A hybrid wavelet decomposer and GMDH-ELM ensemble model for Network function virtualization workload forecasting in cloud computing

Nowadays Network function virtualization (NFV) has drawn immense attention from many cloud providers because of its benefits. NFV enables networks to virtualize node functions such as firewalls, load balancers, and WAN accelerators, conventionally running on dedicated hardware, and instead implements them as virtual software components on standard servers, switches, and storages. In order to provide NFV resources and meet Service Level Agreement (SLA) conditions, minimize energy consumption and utilize physical resources efficiently, resource allocation in the cloud is an essential task. Since network traffic is changing rapidly, an optimized resource allocation strategy should consider resource auto-scaling property for NFV services. In order to scale cloud resources, we should forecast the NFV workload. Existing forecasting methods are providing poor results for highly volatile and fluctuating time series such as cloud workloads. Therefore, we propose a novel hybrid wavelet time series decomposer and GMDH-ELM ensemble method named Wavelet-GMDH-ELM (WGE) for NFV workload forecasting which predicts and ensembles workload in different time-frequency scales. We evaluate the WGE model with three real cloud workload traces to verify its prediction accuracy and compare it with state of the art methods. The results show the proposed method provides better average prediction accuracy. Especially it improves Mean Absolute Percentage Error (MAPE) at least 8% compared to the rival forecasting methods such as support vector regression (SVR) and Long short term memory (LSTM).     

面向5G网络云原生应用资源调度的博弈优化策略

随着5G网络和云原生技术的发展,面向服务的5G云原生核心网应运而生,传统应用正朝着云原生化方向发展。目前云原生服务提供商和云原生应用商数量众多且关系复杂,使得应用在云原生化过程中的资源调度面临新挑战。提出一种5G网络云原生应用资源调度优化策略,将云原生应用商和云原生服务提供商构建为多主多从的Stackelberg博弈模型,对传统收益进行具体描述并联合能耗构建利润函数和策略空间,证明给定一组微服务资源定价的情况下存在云原生应用商的纳什均衡点。在此基础上,引入柯西分布对策略进行优化,提高其收敛性能,通过分布式迭代方法得到云原生服务提供商的最佳微服务定价和云原生应用商的最佳微服务租用比例。仿真结果表明,相比ACA算法、QOS PA算法以及GOS策略,该策略能够有效提高网络收益和用户体验质量,同时降低应用开发能耗。