Seleção distribuída de recursos em grades computacionais usando raciocínio baseado em casos e políticas de granularidade fina

Abstract

One of the main functionalities for virtualization service on grid computing is suitable resource selection for job execution. Resource brokers provide this functionality. Resource selection on current brokers still presents challenges for achieving the best solution in the decision-making process, especially when considering many factors. In our work, the user decides the resource selection in grid computing. We approach this problem considering user preference for specificresource selection objectives such as expected performance for an application execution, resource access restriction, execution application cost, and resource reliability. For each objective, we employ different techniques and combine them in a decision theory model. By consideringperformance in the selection process, we use the case-based reasoning technique based on similar past job executions to predict a new job time execution. In this prediction model, we develop a new case retrieval algorithm for flat memories, which is based on relevant sequence and geometric distance for case attributes. Results show that our prediction model is accurate and efficient in the prediction process. The case retrieval algorithm also presents better performance than otherapproaches as the case base increases. With the resource access restriction as a selection factor, we develop a fine-grain policy-based model for distributed resource access verification. Unlike a globalaccess policy, which applies to all resources in a virtual organization, a fine-grain policy establishes rules for specific resources and users. In this case, a previous access restriction verification prevents a resource selection which may denies access to a requisition, resulting in an unsuccessful submission. In addition, the model developed is based on standard policies that avoid redundancy in access control management. Results show that our distributed model runs faster than centralizedapproaches and presents indexes analyzing the efficiency of each approach based on machines, requisitions, and access restrictions heterogeneities. We consider the resource cost in the selectionprocess as an attribute in service level agreements between resources and users. By considering the resource reliability in the selection process, we use historical data from each execution in the environment, which register the job execution resource probability in the predicted time and with the negotiated cost. The decision model is completely formalized using the multi-attribute utility theory, which relates the important objectives above and allows different proportions of userpreferences for each objective. The complete solution is distributed and implemented using a multiagent system, acting as a resource broker. All models of this thesis are analyzed in a real environment, presenting appropriate functional behaviors and positive performance results.