Development of a hybrid optimization architecture combining static scheduling with dynamic load balancing

Управління розвитком складних систем. 2026. No. 65. ISSN 2219-5300 DOI: 10.32347/2412-9933.2026.65.70-81 UDC 004.4'2:004.942:004.032.26:004.8 Yurii Bilak ORCID: http://orcid.org/0000-0001-5989-1643 Uzhhorod National University, Uzhhorod, Ukraine PhD (Phys. & Math.), Associate Professor, Department of Software Systems Fedir Saibert ORCID: http://orcid.org/0009-0004-8081-4174 Uzhhorod National University, Uzhhorod, Ukraine Assistant, Department of Software Systems Antonina Reblian ORCID: http://orcid.org/0000-0002-2875-2197 Uzhhorod National University, Uzhhorod, Ukraine Lecturer, Department of Software Systems Article history: Received: 20.01.2026 Accepted: 02.02.2026 Published: 26.03.2026 DEVELOPMENT OF A HYBRID OPTIMIZATION ARCHITECTURE COMBINING STATIC SCHEDULING WITH DYNAMIC LOAD BALANCING Abstract. Modern parallel and distributed computing systems are characterized by high complexity, heterogeneous resources, and variable workload intensity, which complicates efficient scheduling and utilization of computational resources. Traditional approaches based solely on static scheduling or purely dynamic load balancing often fail to provide sufficient adaptability and performance predictability under real operating conditions. In this context, developing hybrid methods that integrate preliminary task analysis with adaptive resource management during execution becomes essential. This paper presents a hybrid architecture for optimizing computational processes in parallel environments, combining structural, statistical, and intelligent levels of analysis and integrating static scheduling with dynamic load balancing. The computational process is formalized as a directed acyclic graph of tasks, accounting for resource requirements and inter-task dependencies. Offline profiling allows the construction of empirical models of scalability and execution time, which are then used to form an initial resource allocation. The intelligent component of the architecture is implemented as a neural network module that predicts task execution efficiency depending on placement parameters and the current system state. Regression analysis, multilayer perceptron modeling, and statistical validation with experimental profiling data are employed for model training and assessment. Task allocation optimization is performed using greedy heuristics and genetic algorithms, enabling the identification of compromise solutions balancing execution time and energy efficiency under limited resource constraints. Numerical experiments on block matrix multiplication and two-dimensional heat conduction problems confirm the effectiveness of the proposed hybrid approach. Combining offline analysis with online adaptation significantly improves parallel computing performance and reduces losses caused by load imbalance. Achieved results include speedups of up to 18×, waiting time reduction of up to 27%, and performance prediction accuracy ranging from 93% to 96%. The developed hybrid architecture demonstrates a robust and flexible approach for optimizing computational processes in dynamic heterogeneous environments. It provides a practical foundation for the further development of intelligent management systems in parallel computing, offering predictable performance, efficient resource utilization, and enhanced adaptability to changing workloads. Keywords: parallel computing; optimization; performance models; distributed systems; dynamic balancing; forecasting Introduction Modern computing tasks require ever greater productivity, which necessitates the use of parallel and distributed technologies. At the same time, the need for effective means of optimizing computing processes is 70 growing. Traditional methods have exhausted their capabilities, and today adaptive, intelligent and hybrid optimization methods are relevant. The scientific novelty of the work lies in the integration of three levels of analysis – structural, statistical and intelligent – within a single optimization © 2026 Y. Bilak, F. Saibert, A. Reblian. This article is published under the CC BY-NC-ND license. Інформаційні технології управління architecture. A non-standard use of a neural network is proposed not only as a means of classification or approximation, but also as a predictor of task performance efficiency, which allows for a reasonable allocation of resources. An important feature is the combination of offline and online methods, which provide dynamic adaptation of parameters without the need for a complete restart of the computing process. In addition, the architecture provides flexibility in choosing optimization mechanisms – between heuristic and genetic algorithms – depending on the nature of the task and system constraints. Research goal and objectives. The aim of the work is to develop a hybrid intelligent architecture for optimizing computational processes in a parallel environment, which combines structural, statistical and intellectual levels of analysis, in order to ensure efficient, adaptive and energy-balanced distribution of computational tasks in heterogeneous computing systems. Research objectives 1. Formalize the structure of the computational process in the form of a directed acyclic graph of tasks, determine critical paths and resource intensity of system elements. 2. Build statistical performance models based on task profiling to assess scalability, execution time and energy consumption in different configurations. 3. Develop a neural network forecasting module capable of predicting the efficiency of task execution depending on the distribution parameters and contextual characteristics. 4. Formulate and implement a task distribution optimization problem that takes into account resource constraints and the objective function (time, energy minimization or a combined criterion). 5. Integrate offline and online analysis to build an adaptive architecture capable of dynamically adjusting resource distribution in real time. 6. Conduct numerical experiments to evaluate the efficiency of the developed system using the example of typical tasks with different load characteristics. Literature review. Over the past decades, the optimization of parallel computing processes has been based on classical formal models that laid the theoretical foundation for algorithm design and performance analysis. In particular, the PRAM model (Parallel Random Access Machine) [1] allowed us to consider parallelism in its idealized form – without taking into account delays and resource constraints. However, its main drawback is excessive abstraction, which makes it unsuitable for predicting performance in real systems. The BSP (Bulk Synchronous Parallel) model [2], which takes into account synchronization and communications, reflects the behavior of clusters much more accurately, but it also assumes fixed superstages, which is poorly adapted to dynamic loads. The LogP model [3], which details delays, processing and throughput, provides a bet (...truncated)