Research On Elastic Control For Multi-Tenant Data Storage System In Cloud

The emerge of cloud computing provides a new delivery model for large Internet application, with the characteristics of large scale, high scalability, on-demand service bring new vision for the next generation of application development. The cloud "limit-less" resources and pay-per-use price provide the application with taking full advantage of elastic resource of cloud computing to deal with the load function; this can achieve scalability with respect to traditional applications. Elasticity in cloud is that the ability to add more cloud resource provision in peak in order to meet performance require-ments and release extra resources to reduce overall overhead, which is one of the major characteristics. Due to the increase of scale and complexity of cloud large applications, these applications usually are designed in multi-layer architecture, and now the most of applications are data-driven, the expansion of the data layer is most important to the ex-tension of the whole system. Compared with traditional relational database, there are many different aspects considering the key characteristic of data scalability in cloud computing environment. Due to the inherent statefulness of data storage and requiring an exclusive server, data management should consider distributed transaction characte-ristics, load balancing adaptive problems and copy the consistency strategy, which com-plicate the elastic control of data storage level. SaaS, with the main character of multi-tenant, is the most common type of cloud services, which is recognized be more and more manufactures. The two major characteristics are:single instance and multi-tenant and system can adapt to multi-tenant requirement of elasticity. Multi-tenant feature complicates elastic control of data storage system in cloud, facing the following ques-tions:1) First, it is difficult to satisfy different performance requirements under the con-dition of data nodes shared storing different tenants’data item. For cloud providers, re- source consolidation via shared storage in distribution system is a straight-forward way to minimize the overall resource costs. Consequently, in a shared cloud environment, many data nodes shared store different tenants’data items. However, different tenants have diverse performance requirements described in SLAs, signed between each user and provider, which uses data request latency to express. It increases the difficulty of data elasticity to ensure the above conditions in data elasticity.2) In order to realize adaptive elasticity, data layer need control strategy. The strat-egy decides which data to be moved, where to move. The large application has large load fluctuation; elastic load balancing requirements need scale-out in peak load and scale-down in load low. This frequently data move (for example, partitioned or coa-lesced) could bring new performance problem on overloaded system.3) A load prediction mechanism enables the data scale control strategy in advance. The cloud resource supply need startup delay, it is required to make control decision as soon as possible. Monitoring and control the performance requirements designed in SLA is easily affected by environmental noisy. It is difficult to measure accurately, and the performance requirements of difficult tenants are more complicated with respect to the same requirement.To address above problems, we put forward an AdaptScala system based on MPC control system. This system can monitor different multi-tenant data request workload and make prediction based on single exponential smoothing; given different tenants per-formance requirements on request latency described in SLA, construct performance model in order to decide whether every server can meet different tenants performance requirements on request latency; finally propose an algorithm on data elastic control, which is used to calculate data storage adjustment strategy, the goal is to make the data control little impact on overall system performance, and to minimize the overall cost of resources.In this paper, we use Berkeley DB as key/value data storage engine in cloud, us-ing open source software to build AdaptScala system, simulation multi-tenant data re- quest and collect performance data to build multi-tenant data storage model. Through many experiments demonstrate this system can satisfy proposed elastic control targets.