基于SLA的多租户数据服务利益最大化问题研究

发布时间：2018-05-05 01:18

本文选题：云计算 + SLA　；参考：《山东大学》2016年硕士论文

【摘要】：SaaS(软件即服务)以其“单实例多租赁”的特点为越来越多的用户所接受,在这种服务模式下,一个应用实例可以为多个租户提供服务,减少了服务提供商为每个租户维护独立应用实例的服务成本,并且应用具有十分灵活的扩展性。但是,相对一般单租户的应用,多租户应用中每个租户可以定制个性化需求,其业务特性变得复杂,这就使得不同的租户对于数据库服务性能有着不同的要求,包括服务吞吐量、平均响应时间等。为确保SaaS应用的数据库服务质量,每个租户会与SaaS数据库服务提供商之间协商签订服务水平协议(Service-Level Agreement, SLA)。参照双方的SLA协议约束,SaaS数据库服务提供商提供的服务想要获得收益,必须能使租户数据库作业的响应时间等性能满足协议标准,否则受到相应的违约处罚。因此,SaaS数据库服务提供商需要进行资源的均衡配置和租户作业执行顺序的制定,以满足租户个性化性能需求,另外,在应用实例的运行过程中服务提供商需要进行集群资源监测以及性能优化,保证尽可能为完成租户作业的提供充足的资源环境。云计算环境下SaaS服务质量的控制需要综合考虑多方面的因素。首先,多租户应用通常具有多热点、多混合的负载特征。多热点主要体现在不同租户的业务热点和峰值不同,热点表示租户的业务高峰期和持续时间,峰值表示租户作业最大的负载压力,不同租户的热点和峰值也是相互独立的；多混合体现在不同租户具有不同的负载特征,并且混合在一起会产生更加复杂的负载压力。这二者使得SaaS数据库服务提供商预测租户负载压力和构建作业调度模型变的复杂。其次,租户因为业务特性的不同对于10、MEM和CPU资源的需求程度不一样,有些租户业务偏重计算会消耗较多的CPU资源,有些租户的业务需要进行表的大量关联并缓存大量数据,还有一些租户业务频繁读写硬盘,并且,多租户应用中多个租户的总体负载并不是单纯的每个租户负载压力的叠加,这也加重了对租户负载压力的预测和对节点资源性能承受能力的判断的困难程度。然后,租户的数据在数据库系统中共享存储资源,各个租户在相同数据节点上执行作业时候就会产生资源竞争,一般情况下租户的复杂的数据库作业响应时间长,简单的数据库作业响应时间短,不合理的执行顺序会使得某些租户的长作业长时间占用数据节点的资源引起其他租户短作业违反SLA约束。因此,如何利用租户数据组织的特点,合理均衡租户的负载压力,为每个数据节点中租户的作业制定有效的执行顺序,是众多数据库服务提供商实现总体SLA利益最大化的目的的有效途径。本文对多租户应用的分布式数据库系统里的作业处理进行了系统研究,提出了一个基于无中心数据架构的两阶段作业调度框架AdaptiveSLA,在分布式数据库既有数据放置的情况下,对各数据节点接收的应用服务器发送过来的租户作业进行两阶段调度,实现数据服务提供商SLA整体利益最大化,并尽可能减少执行作业过程中移动数据的代价。首先,在系统初始化阶段,本文使用机器学习算法构建数据节点上的租户性能需求预测模型。然后,调度的第一阶段是数据节点在接收应用服务器发送的数据库作业的过程中,周期性使用危机探测器进行节点的危机探测,如果危机探测器探测到节点发生了性能危机,数据节点使用滑动窗口协议为后续到达的租户作业找到最合适的调度目标以缓解节点的性能危机；最后,调度的第二阶段是在数据节点在执行租户的数据库作业的时候,综合考虑作业SLA约束的响应时间和执行时间为作业制定执行次序,使尽可能多的作业能在规定的时间完成,增加服务提供商的收益减少违约赔偿,以此达到数据服务提供商整体SLA利益最大化的目的。并且,为验证论文方法的有效性和可行性,本文设计了多组实验评估了使用的方法和模型,证明论文方法有效的达到了数据服务提供商SLA整体收益提升的目标。
[Abstract]:SaaS (software as a service) is accepted by more and more users with its "single instance multi tenancy" feature. In this service mode, an application instance can provide services for multiple tenants, reducing service providers' service costs for each tenant to maintain an independent application instance, and the application has a very flexible scalability. Compared with the general single tenant, each tenant can customize the personalized requirement in the multi tenant application, and its business characteristics become complicated. This makes the different tenants have different requirements for the performance of the database service, including the service throughput, the average response time, etc. to ensure the quality of the database service of the SaaS application, each tenant will The service level protocol (Service-Level Agreement, SLA) is negotiated with the SaaS database service provider. According to the SLA protocol constraints of both parties, the services provided by the SaaS database service provider want to gain the benefit, and the response time of the tenant database operation must be able to meet the protocol standards, otherwise the corresponding breach of contract will be available. Therefore, the SaaS database service provider needs a balanced allocation of resources and the formulation of the execution order of the tenant operation to meet the tenant's personalized performance requirements. In addition, the service provider needs to monitor and optimize the performance of the cluster resources during the operation of the application instance to ensure that the tenant is completed as much as possible. The control of the quality of the SaaS service in the cloud computing environment requires a comprehensive consideration of many factors. First, multi tenant applications usually have multi hot and mixed load characteristics. The hot spots are mainly reflected in the different tenants' business hot spots and peak values. The hot spots indicate the peak and duration of the tenants' business, the peak expression The maximum load pressure of the tenant job, the different tenants' hot spots and peaks are also independent; multi - mix embodies the different tenants with different load characteristics and is mixed together to produce more complex load pressure. These two make the SaaS database service provider pretest the tenant load pressure and build the job scheduling model Secondly, the tenants are different for 10, MEM and CPU resources because of the different business characteristics. Some tenant businesses tend to calculate more CPU resources, some tenants need to carry out a large number of associations and cache a large number of data, and some tenant services frequently read and write hard disks, and many tenant applications are used. The overall load of more than one tenant is not a simple overlay of each tenant load pressure, which also aggravates the predictor of the tenant load pressure and the difficulty of judging the capability of the node resource performance. Then, the tenant's data shares the storage resources in the database system, and the tenants perform the job on the same data node. In general, the complex database job response time is long, the simple database job response time is short, the unreasonable execution order will cause the long operation of some tenants to occupy the data nodes for a long time to cause other tenant short jobs to violate the SLA constraints. According to the characteristics of the organization, it is an effective way for many database service providers to achieve the goal of maximizing the overall SLA benefits for each database service provider. This paper makes a systematic study on the job handling in the distributed data base system for multi tenant applications. A two stage job scheduling framework AdaptiveSLA based on no center data architecture is proposed. In the case of distributed data storage, the tenant operation sent by the application server received by each data node is scheduled for two stages, and the overall benefit of the data service provider SLA is maximized and reduced as much as possible. Firstly, in the initialization stage of the system, this paper uses machine learning algorithm to construct the tenant performance requirement prediction model on the data node. Then, the first phase of the scheduling is that the data node is used in the process of receiving the database sent by the application server, and the periodic use of the crisis detector is made. If the crisis detector detects that the node has a performance crisis, the data node uses a sliding window protocol to find the most appropriate scheduling target for the subsequent tenant operation to alleviate the performance crisis of the node. Finally, the second phase of the scheduling is when the node is executing the database operation of the tenant. At the same time, considering the response time and execution time of the job SLA constraints to work out the execution order for the job, so that as many jobs can be completed at the specified time, increase the revenue of the service provider to reduce the default compensation, so as to achieve the purpose of maximizing the overall SLA benefit of the data service provider, and to verify the validity of the method of the paper. And feasibility, this paper designs several groups of experiments to evaluate the methods and models used, and proves that the method of the paper effectively achieves the goal of improving the overall revenue of the data service provider SLA.

【学位授予单位】：山东大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP393.09

【相似文献】