Advantages of Dataflow Computing Model

 
Levchenko N.N., Okunev A.S., Stempkovsky A.L. (IPPM RAS)
 
Abstract - The main development trends in the field of high-performance computing demonstrate the growth of the number of computational cores to hundreds of thousands and even millions. At the same time, a high level of general concurrency of program execution should be maintained. To solve these issues, research and development of new architectures with the use of non-traditional computing models are conducted. This article describes the architecture of the parallel dataflow computing system (PDCS) that implements the original dataflow computing model with a dynamically formed context which is embodied in the parallel programming language DFL. The basic principles of the computing model are described: the operation of program nodes, their activation by data readiness; ways of token formation; the operation of the content addressable memory of keys; the distribution of computations using hashing, etc. The article also describes the main advantages of the computing model and the architecture of the PDCS, which allow creating well-scalable parallel programs for execution on hundreds of thousands of computational cores. First of all, it is the simplicity of creating parallel programs, thanks to which it is possible (for this architecture) to create and debug a program on one computational core, and then the definitive program will be automatically parallelized to any number of cores. The work with memory at the level of interaction between software nodes is "hidden" from the programmer; there is no problem of cache coherence. The PDCS architecture allows at runtime the identification of the task parallelism, which is initially present in the program and is implicit for the programmer. The use of the "scattering" paradigm provides a more efficient data migration strategy. By using hash functions, the programmer can adjust the degree of parallelism of the task execution and the degree of its locality, thereby reducing the loading of the communication network of the computing system. The features of the dataflow computing model presented in the article and their implementation in the architecture of the computing system make it possible to solve the problem of efficient parallelization of tasks on a large number of computational cores, which indicates the prospects of using this computing model for creating supercomputer systems with peak performance.

Keywords - dataflow computing model, architecture, advanced computing, computation scaling.

Преимущества потоковой модели вычислений

 
Левченко Н.Н., Окунев А.С., Стемпковский А.Л. (Институт проблем проектирования в микроэлектронике РАН, г. Москва)
 
Аннотация - Одной из основных задач в области высокопроизводительных вычислений является решение проблемы эффективного распараллеливания вычислений. В данной статье описывается архитектура вычислительной системы, реализующая оригинальную потоковую модель вычислений с динамически формируемым контекстом. Представлены основные принципы работы этой модели вычислений. Проведен анализ особенностей и преимуществ потоковой модели вычислений. Сделаны выводы о перспективности использования потоковой модели вычислений для создания в будущем новых высокопроизводительных вычислительных систем.

Ключевые слова - потоковая модель вычислений, архитектура, перспективные вычислительные системы, масштабирование вычислений.