Automated Optimal Configuration Determination of the Parallel Dataflow Computing System for Solving a Specific Problem

Stempkovsky, A.L.; Bobkov, S.G.; Zmejev, D.N.; Levchenko, N.N.; Klimov, A.V.

Home Authors Papers Year of conference Themes Organizations To MES conference

Automated Optimal Configuration Determination of the Parallel Dataflow Computing System for Solving a Specific Problem

Authors

Date of publication

2021

DOI

10.31114/2078-7707-2021-3-82-93

Abstract

One of the most important trends in modern multiprocessor computing is the increase in energy efficiency of computing, which is closely related to the high real performance of computing systems. One of the reasons for the low real performance is the small number of parameters (in traditional multiprocessor systems) provided to the user, adjusting which he can adapt the system to the requirements of his specific task. The article is devoted to the problem of determining the optimal configuration of a universal parallel dataflow computing system (PDCS) for solving a specific task. The article briefly describes the dataflow computation model and the original PDCS architecture that implements this model. The criteria by which it is proposed to evaluate the PDCS architecture are given, as well as the parameters and their influence on these criteria are described. It is proposed to consider the parameters as global, the change of which affects the architecture of the computing system as a whole, and as local, which affect the behavior of individual nodes and blocks of the system. The global parameters are: the number of computational modules, the number of matching processors, the number of execution units, the type of distribution function of computations, and others. The local parameters are: the number of functional blocks in the execution unit, the memory size of the matching processor, the number of functional blocks in the matching processor, and others. The criteria are: the task execution time, the average workload of the execution units, the cost of the system, the uniformity of data distribution, and the maximum load of the memory of keys of the matching processor. An algorithm for the automated determination of the optimal configura-tion of a computing system for solving a specific task is presented. The algorithm is based on the described parameters and criteria, as well as the use of a behavioral block-register model. Plans for further improvement of the algorithm are outlined. The implementation of this algorithm in the software package for the design of dataflow computing systems will significantly reduce the total research time of the PDCS, as well as ensure the operational transformation of the basic PDCS architecture into a special computer for solving a specific task or a group of tasks.

Keywords

parallel dataflow computing system, optimization parameters, performance criteria, dataflow computational model, optimal architecture.

Library reference

Stempkovsky A.L., Bobkov S.G., Zmejev D.N., Levchenko N.N., Klimov A.V. Automated Optimal Configuration Determination of the Parallel Dataflow Computing System for Solving a Specific Problem // Problems of Perspective Micro- and Nanoelectronic Systems Development - 2021. Issue 3. P. 82-93. doi:10.31114/2078-7707-2021-3-82-93

URL of paper

http://www.mes-conference.ru/data/year2021/pdf/D070.pdf