An Artificial Intelligence Framework for Connectivity Optimization and Adaptive Routing in Processor Interconnection Networks
Abstract Interconnection networks determine the communication efficiency, scalability, and fault resilience of parallel and distributed computing systems. Classical topologies such as Hypercube, Torus, Fat-Tree, Dragonfly, and Perfect Difference Network (PDN) provide different trade-offs among degree, diameter, path diversity, and communication cost, while static routing may perform poorly under dynamic traffic and failures. This paper presents AICOF, an AI-based Connectivity Optimization Framework that integrates Multi-Layer Perceptron (MLP) traffic estimation, Graph Neural Network (GNN) topology-aware learning, and Proximal Policy Optimization (PPO)-based reinforcement learning for adaptive routing. The network is modeled as a graph, and routing optimization jointly considers connectivity, throughput, reliability, latency, congestion, and energy consumption. A unified simulation methodology is applied to Hypercube, Torus, Fat-Tree, Dragonfly, and PDN topologies with network sizes from 64 to 512 nodes and multiple traffic and failure scenarios. The evaluation includes topology comparison, scalability analysis, AI-assisted routing, fault tolerance, and ablation studies. The results indicate that AICOF improves throughput, latency, packet delivery, energy efficiency, and fault recovery compared with conventional and standalone AI-based routing approaches.
Authors
- Sushil Kumar Dwivedi (ORCID: https://orcid.org/0000-0003-2744-9521)
- Rakesh Kumar Katare (ORCID: https://orcid.org/0009-0002-4981-4972)
Institutions
- Awadhesh Pratap Singh University (IN)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-12
- DOI
- https://doi.org/10.21203/rs.3.rs-11260644/v1
- Primary Topic
- Interconnection Networks and Systems
- Type
- preprint