Reconfiguração elástica de componentes paralelos em nuvens de serviços de Computação de Alto Desempenho

Abstract

Researchers willing to run scientific applications have a vast choice of computing infrastructures such as supercomputers and clouds. Since these infrastructures are of shared nature, it is usual to face variability in the state of resources during the execution of a parallel program. End users and developers should have the capability to adapt and control execution to ensure compliance with quality of service requirements. In the particular case of clouds, the reconfiguration is enabled by the concept of elasticity. In this Thesis, we present an architecture for a framework that allows the stakeholders to define policies and mechanisms of control at runtime for the elasticity of components of parallel systems, changing the set of allocated resources. We use the HPC Shelf cloud of components as the execution environment. This cyberinfrastructure presents a contextual contract system that allows the proper description and allocation of components and virtual platforms, defining the optimized match according to the characteristics of the parallel programs and the resources available on the platforms. Contextual contracts also allow the definition of quality of service contracts according to researchers' priorities and requirements. The resulting HPC Shelf environment with elastic reconfiguration control expands the options for the stakeholders, ensuring that the runtime respects the component service quality contract in the face of incomplete information and fluctuations in the environment.