BIOINSPIRED SIMULATION METHOD FOR SCHEDULING OF PARALLEL FLOWS APPLICATIONS IN GRID-SYSTEMS

Abstract

The article is devoted to solving the problem of parallel requests scheduling flows in spatially
distributed computing systems. The relevance of the task is justified by a significant increase in
the demand for the distributed computing paradigm in the conditions of information overflow and
uncertainty. The article discusses the problems of scheduling user requests that require severalprocessors at the same time, which goes beyond the classical theory of schedules. The aspects of
the efficiency of using heuristic algorithms for scheduling planar resources are analyzed. The
reasons for their insufficiency are determined both in terms of effectiveness and empirical approaches.
The paper proposes to solve the problem of scheduling parallel applications based on
the integrated application of intelligent agents coalition and an event simulation model. It is proposed
to classify incoming applications on the basis of using a modified bio-inspired optimization
method for cuckoo search. The joint use of a coalition of intelligent agents and a bio-inspired
method will allow for unprecedented parallelism of calculations, and the subsequent determination
of the processing classified applications ways on the basis of a simulation model will allow us
to form sets of alternative solutions to speed up problem solving and optimize the distribution of
available computing resources depending on the sets of incoming applications. To evaluate the
effectiveness of the proposed approach, a software product was developed and experiments were
conducted with a different number of incoming applications. Each incoming application has a
certain set of attributes, which is a vector of the application characteristics. The degree of the
application similarity feature vector and the vertex reference feature vector in the distributing
simulation model is a classification criterion for the application. To improve the quality of the dispatch
process, new procedures for duplicating unclassified applications have been introduced, which
allow intensifying the search for matches in feature vectors. It also provides backup dispatching trajectories
necessary for processing precedents for the appearance of applications with absolute priority
at the inputs. The quantitative estimates obtained demonstrate time savings in solving problems of
relatively large dimension (from 500,000 vertices) of at least 10%. The time complexity in the considered
examples was O (n 2). The described studies have a high level of theoretical and practical significance
and are directly related to the solution of classical problems of artificial intelligence
aimed at finding hidden dependencies and patterns on a large set of big data.