The present paper is focused on application of modern computing technologies and methods of parallel programming directed to increasing the efficiency of solution of fundamental scientific and application problems. Technological aspects of parallelization of the computing system
A technique for mapping parallel program structures onto structures of robust distributed computer systems (CS) is proposed. Effective algorithms are developed for implementation of the technique stages: 1) a heuristic algorithm for mapping nodes of the parallel program graph onto the graph of the distributed CS; the algorithm substantially reduces the mapping time with respect to the well-known Bokhari algorithm; 2) a decent- ralized algorithm for mapping the program graph edges which do not coincide with edges of the CS graph onto shortest paths on the CS graph. Mapping one-dimensional (line, ring) and two-dimensional (mesh, torus) parallel program structures onto regular structures (torus, two-dimensional circulant, and hypercube) of robust computer systems with faulty components (computers and intercomputer connections) is investigated. It is shown that: one-dimensional parallel program structures are mapped onto structures of distributed CS better than two-dimensional ones; and if defects (failures of the CS components) arise in the CS structure, the quality of mapping one-dimensional structures deteriorates less than the quality of mapping two-dimensional structures.
The method of domain decomposition using a three-dimensional analogue of the Peaceman-Rachford algorithm is considered. Estimations of parallelization efficiency and results of numerical experiments on computers RM-600-E30 and MVS-1000 for different grid domains and processor topologies are presented.
A parallel program debugger GEPARD developed in the Siberian Supercomputer Center is presented. The motives for creating the debugger, and also the architecture, implementation, and applications are considered.
Parallel implementation and computational aspects of the problem of simulating dynamics of gravitating systems, in particular, a protoplanetary disc, are considered. The mathematical model of the disc involves the Vlasov-Liouville multidimensional kinetic equation, the Poisson equation, and gas dynamics equations. One of the main problems of numerical simulation of gravitational dynamics is the necessity of multiple solution of the Poisson equation. A fast parallel solver for this equation and a dynamic load balancing algorithm for processor elements are presented. They employ the physical peculiarities of the problem.
Basic ideas and algorithms for generating parallel programs for numerical modeling of large-scale problems are presented. Generation of a wide spectrum of programs is based on using a good man-made code, an assembly technology of parallel programming, parameterization, and macrogeneration. It is shown that these technological techniques taken altogether ensure a high quality of generated parallel programs and relieve the system user of having to program the procedures of synchronization, dynamic load balancing, and other elements of parallel program development whose implementation is difficult. The approach is implemented in the ParaGen system.
A new approach to image isoline approximation by smoothing splines is proposed. The approach eliminates the stage of initial isoline parametrization and allows one to prescribe
A method is developed and described for fast solving in the spatial domain the problem concerned with initial signal (image) reconstruction by a set of low-resolution images that differ by the mutual coordinate shift. The proposed algorithms and the computational scheme implemented on their basis are constructed in such a way that a signal with the minimal energy (dispersion) is selected from the whole class of digital signals satisfying the set of observed data.
Dependences of the efficiency while recognizing noisy static and moving test objects on the duration of their presentation are experimentally studied. Investigation is conducted for different observation conditions. It is found that with increasing duration of observing the dynamic images of test objects in the presence of a white noise the efficiency decreases at first. It is underlined that upon achieving the minimal value that takes approximately 120 ms it grows and then upon achieving the maximal value taking approximately 400 ms it decreases monotonically. Mathematical modeling of the processes of accumulation of the image signal component in the visual memory is fulfilled.
A method for evaluation of the frequency of impact impulse appearance and qualitative analysis of vibrosignal components is described. It is based on nonstandard wavelet transform. The method is compared with the Fourier transform and its capabilities, advantages, and drawbacks are evaluated. The method is used to analyze a real signal from a test machine with a defective rolling bearing. Analysis of the real signal leads to the conclusion that it is insufficient to consider real signals by analyzing only the Fourier spectrum.
