Reinforcement Learning-Enhanced Multipath Routing Algorithms for High-Performance AI Workload Distribution in Data Centers

View at publisher Summarise with VIRAT
Authors: Selvamani Ramasamy
Cited by: 0
References: 14

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Modern data centers hosting large-scale AI workloads demand network routing mechanisms that dynamically adapt to rapid traffic fluctuations, minimize congestion, and maintain stable performance under oversubscribed conditions. Conventional multipath routing schemes rely on static or weight-based path selection, making them insufficient for AI-intensive environments characterized by bursty communication, heterogeneous workloads, and unpredictable flow patterns. To address these limitations, this study introduces a Reinforcement Learning-Enhanced Multipath Routing (RLE-MR) framework designed to deliver intelligent, adaptive, and congestion-aware routing for highperformance AI workload distribution. The workflow integrates real workload traces from the Google Cluster Data repository, traffic preprocessing, Leaf-Spine network simulation, telemetry-driven state construction, a Deep QNetwork (DQN) routing agent, and Software-Defined Networking (SDN)-based flow enforcement within a closed feedback learning loop. Real-time link utilization, delay, and congestion metrics are continuously monitored to compute rewards and refine the routing policy. Experimental evaluation conducted in an NS-3 Leaf-Spine topology demonstrates that RLE-MR significantly outperforms traditional and learningbased baselines. The proposed model maintains over 97 % throughput at moderate loads and sustains above 90% throughput even at full saturation. Latency remains consistently low (2.2-2.8 ms), packet loss stays below 1 %, and the Jain's Fairness Index remains above 0.93 across all load levels. Overall, the RLE-MR framework transforms static routing into a self-optimizing, telemetry-driven system capable of meeting the stringent communication demands of next-generation AI data centers.

Fields of Study & Subjects

Network Traffic and Congestion Control Software-Defined Networks and 5G Cloud Computing and Resource Management

Publication details

Journal / source
International Conference on Communication, Computing and Electronics Systems (ICCCES)
Pages
1637–1644
Publication date
21 January 2026
Language
EN