Contribuição para o desenvolvimento de uma arquitetura de computação própria ao paradigma orientado a notificações

Abstract

Current software development processes lack techniques for the productive and quality design of software that makes efficient use of the parallel execution capabilities provided by the hardware of the modern computing systems. In this context, the Notification Oriented Paradigm (NOP) has been recently developed. The essence of this paradigm is a new organization for software logic based on precise notifications among causal-logical entities. One of the main effects of this paradigm is to enable smaller coupling in the software structure. As a consequence, NOP allows the exploitation of parallelization and/or distribution in a simpler and more efficient way than more commonly used programming paradigms, such as the Imperative Paradigm and the Declarative Paradigm. However, the dynamics of execution under NOP, based on notifications, is not efficiently performed by the hardware of current computing systems, which are fundamentally based on the von Neumann/Turing (and similar) model of sequential execution. In order to address this drawback, this thesis presents a contribution to development of a computing architecture, named ARQPON, which is suitable for execution of software developed according to NOP computing model. ARQPON was designed based on principles of generality, flexibility and scalability, which allow it to execute NOP software of any level of complexity by fetching it from a program memory, without the need for reconfiguration of the hardware that implements the ARQPON itself. The developed architecture is organized as a fine grain multiprocessor that executes instruction hierarchically into sets of specialized processor cores. Experiments performed on this architecture show that ARQPON presents effective performance improvements in comparative evaluations, under some conditions. In addition, it is demonstrated that some conceptual and implementation aspects of ARQPON, such as the memory access model and the scale of paralelization, can be further improved in order to favor ARQPON in relation to von Neumann implementations and make it an even more attractive alternative for parallel computing applications.