Hard architecture of traditional multiprocessor systems does not provide proportional growth of performance when number of processors is increased. Besides, system performance may even reduce because of system loss rising in parallel computational process organization. That is why modern cluster supercomputers provide on a majority of tasks real performance not more than 10-15% of declared peak one.
Multiprocessor systems with programmable architecture designed in SRI MCS of Southern federal university on basis of original conception, suggested by Russian scientists, allow to reconfigure system architecture under the solving applied task structure without change configuration of current technique. Also these systems provide high real performance and in fact linear growth of performance in proportion to used hardware resources (number of processors in the system).
Unlike most reconfigurable computer systems multiprocessor systems with programmable architecture realize structural procedural method of parallel calculations organization. According to this method parallel program is a sequence of computational structures calls, where each structure is adequate to all-in-one task fragment. Firmware tools provide ability of architecture dynamic reconfiguring according to specified computational structure.
Complex of designed methods of transformation applied tasks into structural procedural form and hardware and software tools allow to use hardware resources more effectively and provide real performance 50%-90% from the system peak performance for wide range of tasks of various problem areas. According to effectiveness of applied tasks solving, cost and mass and dimensions characteristics multiprocessor system with programmable architecture is more preferable than rival multiprocessor systems with hard architecture.
Multiprocessor system with programmable architecture is designed on basis of modular scalability principle. Base unit of multiprocessor system is base module – multilayer printed circuit board. Scalability of system computational capability may be increased by connecting several base modules. System software allows to cluster base modules for solving several non-overlapping tasks or to use all available base modules of the system for solving task of high computational complexity.
Modern developments
Circuit board of computational module Nekkar