A abstração de contratos contextuais para alocação de recursos de sistemas de computação paralela orientados a componentes em nuvens

Abstract

High performance computing has been employing techniques and tools to build solutions for performing tasks that demand high performance in both programs and execution platforms. In the evolution of these techniques, the use of high performance computational clouds is still a challenge, since several users of this area are being reluctant to adopt this model, remaining with the execution of their applications in their own computational resource parks. Among the problems reported by these users are restrictions on the choices users can make in their resources, such as the geographic location of execution platforms, the use of specific computational accelerator, and the added overhead of virtualized hardware execution and communication network. In order to promote access to cloud computing environments to meet the requirements of users in this area, as well as to provide access to heterogeneous resources on demand, it was proposed by the HPC group of the Universidade Federal do Ceará, an environment of cloud computing called HPC-Shelf. HPC-Shelf is based on the component orientation paradigm, using component abstraction to represent programs and execution platforms. In this scenario, this thesis introduces Alite, a method for choosing Computer Systems, formed by pairs of components that represent the programs and their respective execution platforms. Alite, through a series of abstractions, is able to treat the universe of components, registered in a database, using its characteristics that can be represented hierarchically, or through numerical values that can be set by users or automatically calculated from the valuations of other numerical characteristics. An Alite implementation prototype was developed with the purpose of evaluating the contextual contracting system with respect to aspects of expressiveness, performance and effectiveness, evidencing its applicability in real scenarios. Regarding expressiveness, a virtual platform framework was designed for HPC Shelf, which constitute heterogeneous cluster computing platforms with some features considered common in modern cluster design. From this framework, a virtual platform framework on Amazon’s EC2 service is presented. Finally, a matrix multiplication computing component framework was developed based on the BLAS library’s level 3 interface, demonstrating how features of widespread scientific libraries can encapsulated and reused through components at HPC Shelf. User contract resolution time was considered satisfactory for a large number of registered platforms, as well as the result of the classification was aligned with the objectives defined by HPC-Shelf.