Allocation de ressources élastique pour l’optimisation de requêtes

Abstract

Cloud computing has become a widely used way to query databases. Today’s cloud providers offer a variety of services implemented on parallel architectures. Performance targets and possible penalties in case of violation are established in advance in a contract called Service-Level Agreement (SLA). The provider’s goal is to maximize its benefit while respecting the needs of tenants. Before the birth of cloud systems, several studies considered the problem of resource allocation for database querying in parallel architectures. The execution plan for each query is a graph of dependent tasks. The expression "Resource allocation" in this context often implies the placement of tasks within available resources and also their scheduling that takes into account dependencies between tasks. The main goal was to minimize query execution time and maximize the use of resources. However, this goal does not necessarily guarantee the best economic benefit for the provider in the cloud. In order to maximize the provider’s benefit and meet the needs of tenants, it is important to include the economic model and SLAs in the resource allocation process. Indeed, the needs of tenants in terms of performance are different, so it would be interesting to allocate resources in a way that favors the most demanding tenants and ensure an acceptable quality of service for the least demanding tenants. In addition, in the cloud the number of assigned resources can increase/decrease according to demand (elasticity) and the monetary cost depends on the number of assigned resources, so it would be interesting to set up a mechanism to automatically choose the right moment to add or remove resources according to the load (auto-scaling). In this thesis, we are interested in designing elastic resource allocation methods for database queries in the cloud. This solution includes : (1) a static two-phase resource allocation method to ensure a good compromise between provider benefit and tenant satisfaction, while ensuring a reasonable allocation cost, (2) an SLA-driven resource reallocation to limit the impact of estimation errors on the benefit and (3) an auto-scaling method based on reinforcement learning that meet the specificities of database queries. In order to evaluate our contributions, we have implemented our methods in a simulated cloud environment and compared them with state-of-the-art methods in terms of monetary cost of the execution of queries as well as the allocation cost.