A TOPOLOGY-AWARE UNIFIED MODEL AND METHOD FOR FORECASTING ENERGY CONSUMPTION IN DISTRIBUTED SYSTEMS WITH STATIONARY AND MOBILE DEVICES

Abstract

The subject of this research is energy consumption modeling and forecasting in distributed computing systems using Entity-Component Architecture (ECA). This task focuses on improving the method of quantifying and predicting the energy demands of a distributed computing system by simulating all components of this system, such as computational units (CPUs, GPUs), data transmission interfaces (LAN, USB interfaces), and network infrastructure (LAN switches, routers) in heterogeneous environments. Method validation is based on the stationary computing network (SCN) and mobile computing network (MCN) cases. This study addresses the inherent limitations of traditional algebraic models by advocating for a modular, scalable, and flexible ECA-based framework that integrates direct and indirect power expenditures of distributed computational elements to facilitate energy profiling. The experimental validation showed that the proposed ECA model achieves high prediction accuracy. In SCN configurations, simulations matched real-world energy measurements within 4.4% for CPU-only tasks, while CPU+GPU tasks showed a larger discrepancy of 20.2% due to GPU underutilization. Conversely, the MCN scenario, involving mobile devices and USB connections, resulted in minimal deviations of 3.5% in CPU-only mode and approximately 10.1% in CPU+GPU mode. These results underscore the model's robustness in accurately forecasting energy usage across diverse computational settings. We conclude that the effective transition from conventional algebraic models to a flexible and extendable ECA approach significantly enhanced the methodology’s modularity, scalability, and universality. Major achievements include refining computational benchmarking practices by generalizing computing payloads, adopting all-inclusive energy measurements, and highlighting the energy-efficiency advantage of computations on mobile platforms over traditional stationary hardware for both CPU and GPU tasks. This research has highlighted some limitations of current methods, such as linear interpolation in load-energy modeling and the neglect of momentary hardware states, including momentary temperature and temperature inertia, which affect performance and consumption. The scientific novelty of the research consists of two parts: a new approach to mathematical modeling that enables extension and scalability, and a more consistent and accurate model of a distributed computing system based on it.