Home         Authors   Papers   Year of conference   Themes   Organizations        To MES conference

The application and implementation issues of dataflow computing system
Authors
	Klimov A.V.
	Levchenko N.N.
	Okunev A.S.
	Stempkovsky A.L.
Date of publication
	2016
Abstract
	Computing model with dataflow management was originally conceived as a more adequate for hardware in comparison with to von-Neumann model. It is much better uses a proper "natural" parallelism of hardware. A classic postulate of distributed programming is: "to store there where is computing" or, in another formulation, "to compute there where to be stored". But then, at the time of use it is inevitable to have a request to data storage with a long feedback waiting time - it is the paradigm of "collection". But the opposite solution is possible: to store data there where they are expected to be used. That is, the computing of data item will immediately initiates its transfer to the place where it will be used. This distinguishes the paradigm of "distribution" from the paradigm of "collection". In terms of inter-processor communication and computation distribution, the dataflow model, which works in the paradigm of «distribution», has the following advantages: 1) Instructions for distribution and modification does not affect the essence of the algorithm; 2) In the paradigm of "distribution" the data transfers in one direction and at once "on destination". To date, the most common approaches to the creation of parallel programs are MPI, Charm++, Polyphonic C#. The MPI, in the light of our classification, should be described as mixed: inside one MPI-process it is a common programming in the paradigm of "collection", on the level of the group of processes - in the paradigm of "distribution". Charm++ and Polyphonic C# are the languages with programming in the paradigm of "distribution", however, their peculiarity, in contrast to our approach, is to follow the object paradigm. The IPPM RAS develops the architecture of the paral-lel dataflow computing system (PDCS), which implements the dataflow computing model with dynamically formed context. For matching of tokens in the system responds the matching processor, element of which is an associative memory of keys. The proposed computing model is distinguished from the classic by the fact that the paradigms of computation and programming were changed. This allows it to be much closer to parallel hardware, more natural way to create parallel programs, to reduce the semantic gap between the high-level parallel language and the proposed architecture in comparison with modern computer systems
Keywords
	paradigm of "distribution", new paradigm of programming, parallel programming, parallel dataflow computing system.
Library reference
	Klimov A.V., Levchenko N.N., Okunev A.S., Stempkovsky A.L. The application and implementation issues of dataflow computing system // Problems of Perspective Micro- and Nanoelectronic Systems Development - 2016. Proceedings / edited by A. Stempkovsky, Moscow, IPPM RAS, 2016. Part 2. P. 100-106.
URL of paper
	http://www.mes-conference.ru/data/year2016/pdf/D114.pdf