Научен доклад ID 2664 : 2025/2

This paper presents a comprehensive experimental study of leader-election mechanisms in dynamic Fog/Edge environments characterized by churn, message loss, and heterogeneous resource conditions. Classical leader-election protocols such as Bully and Ring exhibit limited resilience under rapid node failures, often triggering election storms and producing unstable coordination behavior. To address these limitations, we propose an enhanced approach, SENTRY-L, and its predictive extension, SENTRY-L NPA, which integrate risk-aware assessment, asynchronous authority handover, and neural stability forecasting based on a lightweight GRU model.

The experimental evaluation covers eight churn scenarios with varying message-drop probabilities and cluster sizes. Performance metrics including election time, leader-change count, false re-election rate, unhealthy leader ratio, energy-rank deviation, and communication overhead—were analyzed using boxplot-based statistical visualization. The results demonstrate that SENTRY-L NPA consistently outperforms both the baseline and non-predictive variants, achieving up to 52% faster leader election, 78% fewer re-elections, and an 85–86% reduction in erroneous or unstable leader selections. These improvements confirm that transitioning from reactive to adaptive and predictive leader-election methods substantially enhances stability, robustness, and coordination efficiency in distributed Fog/Edge environments.