Multi-Agent Reinforcement Learning for Coordinated Routing and Traffic Engineering in Distributed Data Center Systems
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 center systems support a wide range of latency sensitive applications, while their routing and traffic engineering mechanisms remain dependent on static or heuristic driven strategies. Traditional approaches such as equal cost path routing and rule-based traffic engineering often fail to respond effectively to rapid fluctuations in workload intensity, leading to queue accumulation, uneven link utilization, and unpredictable latency under high demand. These limitations highlight the need for routing models that can interpret dynamic traffic behavior and adjust decisions in coordinated manner. This study introduces a learning-based routing framework that integrates a Transformer encoder with an Actor-Critic Multi Agent Reinforcement Learning (MARL) architecture. The Transformer encoder extracts temporal dependencies from workload sequences derived from Google Cluster traces, providing each routing agent with a richer state representation. The multi agent framework enables every data center node to act as an autonomous decision maker while still coordinating actions through shared environment feedback. The agents learn routing policies that minimize latency and reduce congestion by anticipating traffic variations. Experimental evaluation demonstrates that the framework maintains consistently low latency between$\mathbf{1. 8 2 m s}$and$\mathbf{2. 2 1 m s}$, achieves a packet delivery ratio of 99.21%, sustains effective goodput of 98.74%, and distributes link utilization more evenly across the network fabric. The policy converges quickly and shows stable decision behavior with minimal oscillation. These findings indicate that combining Transformer based representations with coordinated reinforcement learning provides a promising direction for improving routing efficiency in modern data center environments. The results pave the way for valuable insights into adaptive traffic engineering and intelligent network control.
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 Intelligent Data Communication Technologies and Internet of Things (IDCIoT)
- Pages
- 1018–1025
- Publication date
- 4 February 2026
- Language
- EN