High-performance heterogeneous processing in concentrating computing system

In this paper the possibilities of organizing heterogeneous computing in so-called Combined Architecture systems consisting of a basic host subsystem (a massively parallel computer), and a set of high-performance specialized parallel coprocessors (hardware modules) executing the main workload are discussed. To optimize the choice of hardware modules, a classification of massive computations, based on the notion of processing types, which correspond to the character of data processing is suggested. A technique of parallel programming for the suggested concentrated heterogeneous systems, ensuring close matching of the tasks to the hardware modules is introduced.