Architectural patterns for Parallel Programming : models for performance estimation

Parallel Programming relies on the coordination of computing resources, so that they simultaneously work towards a common objective. Achieving this requires extra effort from the software designer, because of the increased complexity involved. Furthermore, as Parallel Programming is considered a means to improve performance, the software designer has to consider sophisticated and cost-effective practices and techniques for performance measurement and analysis. In particular, it is of great interest to obtain per­ formance information during design stages and before implementation, since this enables the software developer to select the organisation of computations and communications between components. The Architectural Performance Modelling Method is presented as a criteria for selecting the organisation of a parallel program based on estimating its probable per­ formance. By considering a parallel program as an instance of a software architecture, it can be described in terms of interacting software components. Such components can be classified depending on their particular objective and their rate of change, for example, as components associated with the hardware and software environment (or Platform), components representing the fundamental structural organisation for execution and com­ munication (or Coordination), and so on. The performance of a parallel program can be estimated as the result of the contribution of each one of those kinds of components. An Architectural Performance Model is based on selecting from the Architectural Patterns for Parallel Programming (descriptions of coodinations commonly used in Par­ allel Programming), a component simulator (representing a simulation of a processing component’s behaviour), and a performance analysis of parallel applications (in which the information on system performance is examined). Parallel programs simulated using the Architectural Performance Modelling Method range from a complete parallel pro­ gram to a partially implemented program design. The simulation of parallel systems, using the information about the problem to be solved, the available resources, and archi­ tectural patterns describing overall coordinations of the parallel programs, makes it pos­ sible to identify the best performing architectural solution for the system being built.