On the problem of automatic development of parallel applications for reconfigurable computer systems

Рассмотрена оригинальная методика отображения информационного графа прикладной программы на архитектуру реконфигурируемой вычислительной системы с помощью методов редукции производительности, обеспечивающих решение задач, аппаратные затраты на реализацию которых превышают доступный вычислительный ресурс. Доказаны теоремы о свойствах последовательного применения редукций по числу базовых подграфов, по числу вычислительных устройств и разрядности. На основе доказанных теорем и следствий из них сформулирована методика редукционных преобразований информационного графа прикладной программы для автоматической адаптации к архитектуре реконфигурируемой вычислительной системы. Приведена оценка максимального числа преобразований согласно предложенной методике для сбалансированной редукции производительности и аппаратных затрат прикладных программ для реконфигурируемых вычислительных систем. To solve applied problems, the hardware costs of which exceed the available computing resource of FPGA-based computer systems, an original technique was developed for mapping the information graph of an application program to the architecture of a reconfigurable computing system. The proposed technique is based on the performance reduction methods that reduce the productivity of an applied task, which, along with the reducing productivity, does so for the hardware costs of its implementation and, thereby, solve the problem on the available computing resource. We demonstrate that the decrease in hardware costs for the computing structure realization occurs only during the reduction the basic subgraph number, the number of computing devices in a basic subgraph and the data width. The influence of sequential reduction transformations on the computing structure of a problem is examined. The proved theorems are concerned with the possibility of representing the reduction coefficient as a product of the coefficients of successive reductions, on the inability of additive increase in reduction coefficient during sequential reductions and on the superposition commutativity of different sequential reductions. The proved theorems and the corollaries presented in the article allow formulating the basic principles for the method of reduction transformations of the information graph of the problem for adaptation to the architecture of a hybrid reconfigurable computing system. A distinctive feature of the technique is a relatively small number of transformations for a balanced reduction of the information graph of the problem and the implementation of the task on a reconfigurable computer system.The comparatively small number of transformations required for the balanced reduction of the information graph of the problem and for the implementation of calculations on a reconfigurable computer system is the distinctive feature of the technique. For the developed technique, we estimated the maximal number of transformations and found out the decrease in the quantity of analyzed reduction variants from each class. The proposed technique permits the significant reduction of the time needed to create the computational structure of a parallel program adapted to the architecture and configuration of the reconfigurable computing system. Furthermore, the technique allows automatization of this process using the specialized software and providing at least 5075 efficiency in comparison with the solutions of the same problems by specialists.