OPEN DATAFLOW-SYSTEMS WITH NETWORK STRUCTURE

The paper considers open computing systems, which provide the necessary growth of performance and memory by mechanical addition of new units without affecting the existing software environment. Such computing systems are based on the application of a special functionally complete element base (planner, functor, communicator) that implements parallel processing using dataflow control flows when necessary program fragments are transferred along with the data. Aimed at this, when a certain procedure is found ready for starting (the planner is ready for all the data it needs), the corresponding part of the program is opened in the planner – the operator, which is then transferred along with the data to the free execution device – the functor. The result is always returned along the same route by which the procedure was activated. The layouts of the open systems using two units of design are considered: cells on the reduced element base, and servers assembled from these cells. A two-level distributed switching environment is used. At the cell level, it is provided by the transit properties of the planners and functors, and at the server level – by communicators that are part of the cells. Three types of cells are identified, which enable the growth of computing system functions: cells for increasing the number of gateways used to exchange with the external environment, cells for increasing control and working memory, cells for increasing performance. The recession of the concentrated switching environment allowed for these types of computing systems extensions to be performed independently and without any restrictions on their size. A three-dimensional structure of open systems is described, which can be used to build supercomputers.