Abstract

Peer to peer (P2P) overlays dominate content distribution, collaborative applications, and edge services because they eliminate single points of failure and exploit aggregate bandwidth. Yet, heterogeneous node capacity, churn, and route redundancy often throttle end to end throughput. Classical P2P rate control and scheduling schemes (e.g., tit for tat, rarest first) optimise a single objective or operate on a single network layer, leaving cross layer interactions unexploited. This results in sub optimal bandwidth utilisation, especially under bursty traffic and high churn. We introduce E ThruEnsemble, an ensemble algorithm that fuses (i) adaptive chunk scheduling, (ii) topology aware path selection, and (iii) reinforcement learning guided rate control. The three weak learners each make local throughput estimates; a lightweight Bayesian combiner assigns dynamic weights based on recent prediction error. The final scheduling decision maximises a composite utility that jointly rewards link utilisation and delivery deadline satisfaction. We implement the scheme in NS 3 and instrument it with real trace latency variations. In 500 node overlays, E ThruEnsemble raises average throughput by 29?% over BitTorrent’s choking algorithm, 17?% over ML DOS, and 11?% over ChunkyStream while lowering 95 th percentile latency by 22?%. It converges within 25?seconds after a 20?% churn event and achieves a Jain fairness index of 0.93. Sensitivity studies confirm robustness to packet loss rates up to 5?%.

Authors

P. Prabaharan¹, M. Reshma²
Vivekanandha College of Engineering for Women, India¹, University B.D.T. College of Engineering, India²

Keywords

Peer to Peer Networking, Ensemble Learning, Throughput Optimisation, Adaptive Scheduling, Ns 3 Simulation