| With the development of cloud computing and the maturity of application software, Software as a Service (SaaS) as an application type of cloud computing is gradually becoming an important way for small and medium enterprises to apply advanced technology. In SaaS model, mature service providers use single-instance multi-tenant way in general. The same application instance provides service for different tenants, which is called multi-tenant application. Service providers deploy multi-tenant applications on clouds and tenants rent these applications in pay-on-demand way.A tenant rents its SaaS applications with different qualities of service from service provider according to its own demand and capacity to pay. In order to achieve stable quality of service, tenants sign service level agreement with SaaS service providers. Query response time is the greatest concern to users in SLA(Service-Level Agreement). If response time exceeds the required ceiling, users cannot obtain their results of data processing in time, which results in a low SaaS experience. When query response time violates service level objective, the service provider needs to pay a certain penalty for the tenant according to the signed SLA. Therefore, how to do query optimization, improve query efficiency, satisfy SLA with different tenants and minimize SLA penalty cost have emerged as a concern of service providers.Service provider hopes to use less resource cost to satisfy SLAs of all the tenants in the view of cost and profit. Thus multi-tenant database needs to share processing resources among tenants and optimize resource usage. There exist multi-tenant contention of resources and some queries of tenants would violate SLA in multi-tenant query processing with shared resources. In order to minimize SLA penalty cost of service provider, optimizing SaaS multi-tenant data query in cloud computing environment is indispensable. The main issues and challenges are as follows. (1) Multi-tenant data processing needs good cloud infrastructure. Multi-tenant database has the characteristics of massive tenants and huge data volumes, and tenants constantly join and leave the database. It needs to rely on cloud computing platform to complete multi-tenant data processing. Many nodes and large amounts of data require good data organization, node organization and data locating methods, which provide basis for optimizing SLA penalty cost. However, there are little literatures aiming at efficient and clear multi-tenant data cloud infrastructure. (2) the resource allocation granularity of tenant is large and it has the space of further optimization. Allocating resources in units of tenant is easy to realize and the present optimization of SLA penalty cost is in the granularity of tenant. However, different queries of a tenant have various attributes in the aspects of penalty cost, access frequency and the amount of the occupied resources. Thus it is necessary to allocate and schedule resources in units of query in order to do query optimization in fine granularity. (3) There are many tenants and large numbers of concurrent queries in multi-tenant database, which results in performance bottleneck of query processing. Especially when the loads are high, multiple nodes on clouds have unbalanced loads and some queries would not finish before deadline, which increases SLA penalty cost. (4) When processing nodes run at full capacity, many queries would violate the agreement. After each processing node on clouds is configured, the arrival rate of multi-tenant data queries is unstable. When the number of queries are at peak, each query competes to take up the limited resources. At the time adding a new processing node or migrating data cannot quickly address the issue of resource contention. Therefore, we need to design an emergency response mechanism in the peak period in order to minimize the penalty.The thesis aims at optimizing penalty cost of service provider in the cloud computing environment. It combines the characteristics of multi-tenant data shared storage, tenant isolation, application lease customization, and discusses the index, cache and scheduling of data query optimization. Our main work and contributions are as follows.(1) A multi-tenant index mechanism with P2P structure is built. It organizes multi-tenant data, index and nodes on clouds, which avoids performance bottleneck of centralized index and provides a good data organization basis for further SLA query optimization.The index supports the demand of tenant query for isolation, i.e., it avoids to get the invalid data of the other tenants when using index. The mechanism supports the sequential storage of index item and the common compare-style query and range query. It stores the index and data of a tenant on little nodes as possible to avoid large data transfer. Our index also has dynamic scalability and can provide index service for an unlimited number of tenants using the scalability of cloud platform. Extensive experiments show that its time of single node query and range query respectively save 50% and 75% than centralized index, and penalty cost saves 20%when the number of tenants and computing nodes are large.(2) A multi-tenant SLA-aware data cache management mechanism is designed. It optimizes the cache of multi-tenant database according to query characteristics and default penalty cost of different tenants.We construct the quantitative relationship between cached data and query penalty cost, which provide the basis for selecting cached data. It generated cached data for each computing node, which largely reduces the total penalty cost. It also finishes the adjustment of cached data across nodes in a high efficiency. In the mechanism, any computing node can quickly finish the distribution of tenant queries and let a tenant query be processed on the computing node with the shortest operation time. It is experimentally shown that the penalty cost reduce at least 30%than the benchmark algorithm on cloud computing platform.(3) A de-centralized scheduling approach of multi-tenant query to minimize SLA penalty cost is proposed. It determines the processing node and processing time for each query. When the resources are tight, it could make sure that queries are returned before deadline to achieve the minimization of penalty cost.The approach gives each tenant query a priority according to its default penalty cost and deadline urgency. The query with high priority is first processed. It makes each computing node participate into scheduling based P2P structure, which avoids performance bottleneck of scheduling. It improves the data structure of wait queue of tenant query, which helps quickly finish the operations of search, insertion and deletion among massive tenant queries and has a high scheduling efficiency. In the experiments, when the number of tenant queries is large, its penalty cost reduces 50% than benchmark solution. Its time complexity reduces from O(N) to O(log~2 N). Experiments demonstrate that the scheduling time of a tenant query is 2ms or so and it does not change with the increasing number of tenant queries. |
PDF Full Text Request