Coordenação e reconfiguração dinâmica de componentes em uma plataforma de computação paralela

Abstract

Long running applications are very common in High Performance Computing (HPC) domains. During the execution of this kind of application, some improvements or corrections can be identified and applied without making invalid the data that has been processed. In these cases, the ability to make changes in a parallel program during execution is considered useful. Dynamic reconfiguration is the term used to describe this technique, which can be implemented using different alternatives, like architecture description languages (ADL). An ADL allows a the specification of a software based on exogenous conectors to combine and define data types and protocols for orchestrating the interaction between components. This research has the goal of designing an ADL and implementing a configuration interpretation environment for the HPE component-based parallel computing platform. The main purpose of this language is to provide the ability to specify exogenous connectors and support dynamic reconfiguration. Case studies evaluate the performance of the component interpretation developed by the ADL, as well as validate the actions of dynamic reconfiguration. The results are an evidence that the overhead in the interpretation process for realistic problem instances is acceptable, in such a way that, when used wisely, the connectors can be used even in production scenarios. In some cases, the interpretation weight can be disregarded. The reconfiguration experiments are also deemed satisfactory, making the simplicity of the mechanism the major draw of the solution.