The dissertation is devoted to improving the structure and performance of multiprocessor systems with extendible computing segments by aggregating network interface channels adapted to the studied class of problems with the introduction based on the numerical and analytical algorithms and methods, which increases the accuracy and reliability of experimental data processing. The development of theoretical principles and practical implementation of a multiprocessor system with extendible computational segments aims to solve applied problems and increase the productivity and accuracy of mathematical simulation. The thesis laid new and developed existing theoretical and practical foundations of multiprocessor computing systems design with an extendible computing domain based on the aggregation of network interface channels, which allowed increasing the accuracy and reliability of experimental data processing.
The following main scientific results were obtained:
For the first time, a multiprocessor system module with extendible computing segments was developed. Due to the procedure of network interface aggregation, additionally managed switches, switch buffers, redundancy mechanisms of the main components of the module gives increased computing performance for application tasks.
For the first time to identify and establish the features of the network interface of multiprocessor computing systems with expandable computing segments, there was researched the variant of a virtual computer with unlimited memory and a comparative analysis with a real multiprocessor system, that allowed establishing the main factors that affect the entire multiprocessor system performance.
For the first time, based on the analysis of the network interfaces of multiprocessor complexes with extendible computing segments, there were obtained the appropriate analytical ratios for determining performance indicators and analytical ratios that determine the optimal number of nodes of a modular multiprocessor system.
The technology of restructuring the network interface architecture of a multiprocessor system was improved, allowing the identification of the main patterns of computation time for the aggregation mode depending on the expansion of the multiprocessor system's computing domain. Also, the current model of aggregation of network interface channels of multiprocessor systems provides more significant
opportunities for data transfer among computing nodes, significantly improving the features of its performance, speed, and reliability of its operation.
The approach to unlimited parallelism in related problems was further developed by parallelizing the algorithm via direct methods and developing a parallel algorithm on the Cauchy problem's solutions, which in comparison with traditional methods, can
significantly increase such indicators as efficiency and speed of relevant computations. Based on modern parallel computing technologies, a practical approach to solving the limit and coefficient inverse thermal conductivity problems in the extreme formulation
was further developed. It is found that the minimization of the function as the root mean square residual by solving direct problems can be reduced to the problem of separating its minimum and refining this minimum by iterations according to interpolation formulas.
That allowed in comparison with the traditional approach to increase the accuracy and higher efficiency of computations, reduce machining processing time.
The practical significance of the paper is that the proposed multiprocessor system with extendible computational segments aims at solving applied problems and allows increasing the productivity and accuracy of mathematical simulation. The proposed module of a multiprocessor computing system with extendible computing segments should be used as an integrated environment for implementing the process of mathematical modeling of applied problems.
The offered system allows the accelerating process of serviceability restoration of the personal computer multiprocessor system by automated replacement of those modules out of order.
Due to the network interface aggregation, it is possible to significantly increase the performance features of the modular system in different modes of network interface, so for the two-channel network interface system the deceleration coefficient is reduced by 39.5% compared to single-channel and 66.2% compared to four-channel deceleration computing for a four-channel system is reduced by 45.4% compared to a two-channel network interface system;
Developed hardware and software, mathematical models, algorithms, and methods are implemented as industrial designs and software products.
