| Software-as-a-Service, i.e. SaaS, is an important software delivery model of cloud computing with the development of network and maturity of application software and it becomes a significant way of promoting the application of advanced technology by small and medium-sized enterprises gradually. Nowadays the PaaS platform, which can shield most of the bottom implementation of SaaS application developments, gradually becomes the dominant force in cloud computing market growth stage. As the stripping of the application implementation, the SaaS application developers can concentrate on the software and service development, meanwhile, the cloud computing developers may pay all their attentions to the dynamically scheduling of resources and service.At present, most of the SaaS service providers adopt the "single instance multi-tenancy" mode. In this mode, numbers of tenants share a single application instance and the physical data storage but the end users could not feel sharing the application with other users in the same time. The customized multi-tenants sharing storage architecture takes full advantage of resources as hardware and database et.al, and realizes the all levels resources sharing form the operating system to the data structure. However, for the hardware restriction of single data node, we need employ the cloud computing architecture to realize the dynamic propagation of the cloud computing scale and provide more storage space and higher quality of service for tenants.The existing cloud database, such as SQL Azure, provided data leasing services for traditional application and realized the placement of large-scale multi-tenant data, but they did not take the "single instance multi-tenancy" into account in design for multi-tenant application. The multi-tenant application is a special application which can support multiple application level of tenants with many users. Those tenants are managed by the application layer. For application layer all those tenants are isolated with each other, but for data layer all those tenants share the data storage and have no difference with each other. Because feeling unknown of the tenant, the data placement technology existing could not meet the need of multi-tenants application in performance and operating costs, etc.The multi-tenants placement technology which is the foundation of efficient operation for PaaS, mainly demands:(1) Large-scale tenants’ high concurrent transactions requirement. For the instance of Salesforce, thousands of SaaS tenants lease the same application with various data requirements and transaction characteristics. The existing extension on cloud database model as universal table, pivot table, private table etc, could not meet the need of multi-tenant application extension and large-scale high concurrency transaction performance. Meanwhile the tenant data isolation mechanism is difficult to effectively support the data distribution on multiple nodes and redundancy placement.(2) The balance of tenant data distribution. As the growth of the tenants, the data scale increases sharply with the characteristic of multi data center, multi nodes distribution and multi replica placement which leads to the phenomenon of frequent load imbalance for the difference of the tenants’ business scale and complexity. Because of the "single instance multi-tenancy" mode, it is hard to balance all the tenants’ data though the statistics of visit region and access time. To improve the multi-tenants operating efficiency, it is desiderated a distribution policy for tenant perception.(3) The tenant dynamic data migration issues. The customization of tenant data modal, the continuous operation of tenant data and the join and exit of tenant, all those may lead a balanced data node to unbalanced state. For this problem, we have to set up a load testing and dynamic migration mechanism for multi-tenant data modal to ensure that the full use of resources.These problems make a few tenants SLA (Service Level Agreement) cannot be satisfied, meanwhile, it is a waste of the SaaS service provider to provide higher performance for some tenants. This paper focuses on the tenant data placement researches to satisfy the demand of multi-tenants data notes load balance and cost minimization in order to support rapid development and delivery furthest and guarantee efficient operation of the multi-tenant SaaS application. The main contribution of this paper generalizes:(1)We promote a local data migration strategy facing tenant granularity lock and transaction driven to ensure the efficiency of the multi-tenant high concurrent transaction.Considering the majorization of multi-tenant data storage and access model, this paper propose a granularity lock for tenants which make the granularity between ceiling particle size and lower limit of particle size on the average sense for tenants. At the same time, we set up a transaction analysis model combing with transaction conflict rates, deadlocks, the number of lock and the system throughput to realize the transaction driven local data migration strategy and algorithm, providing higher performance for large-scale multi-tenant transaction concurrency and data management. Experiment shows that the model of the tenant’s lock granularity and the migration strategy can increase the response of transaction more than100per second.(2)We give a multi-tenant data copy number solution strategy based on M/M/S/oo model to realize the high reliability, efficient data access of tenants and reduce the cost of service providers.Considering the system load imbalance and the latency of average request caused by tenant disequilibrium data access, this paper studies on the request arrival time and response time which are in accordance with the probability of exponential distribution regularity and determine the optimal multi-tenant data copy number by M/M/S/oo model. It can implement efficient access of data though the replication with multi-tenant data segments on different nodes. Experiment shows that according to the copy solving strategy, the response time of requests and the expected of fitting degree is very high, and can adapt to the changes of the number of requests in system. (3)We put forward a new migration strategy. It could calculate the migration cost to set up the model, reduce the total cost of migrating operation and optimize the overall performance of SaaS application.Because the existing resource allocation model, capacity planning and resource allocation methods could not identify the tenant characteristics, it is hard to resolve the problem of on-demand multi-tenant data migration. For this reason, this paper proposes a new migration strategy, in which demand estimation model determines when to transfer, migration cost model guides to select suitable tenants for migration, and the single write read migration model guarantees the consistency of the transaction in the migration process. This new migration strategy can effectively reduce the total cost of migration operation and optimize the overall performance of the SaaS application. Experiment shows that the implementation of the migration policy can control the impact on events of service requests within10ms.Whereas, the overall utilization rate of the data nodes decreases from5%to15%, the overall resource rate of supplying tenants improves5%to10%. |
PDF Full Text Request