Effective TCP Flow Management Based on Hierarchical Feedback Learning in Complex Data Center Network.

Area of Science:

• Computer Science
• Network Engineering

Background:

• Existing Transmission Control Protocol (TCP) flow control schemes enhance bandwidth utilization between clients and servers.
• However, these methods fail to optimize bandwidth use in complex, hierarchical data center networks due to a lack of path design consideration.

Purpose of the Study:

• To propose a novel TCP flow management scheme for hierarchical complex data center networks.
• The goal is to achieve optimal bandwidth utilization by dynamically managing TCP flow paths.

Main Methods:

• A reinforcement learning-based approach using a hierarchical feedback model is employed.
• This method dynamically controls TCP flow paths to establish an optimal flow policy.

Main Results:

• The proposed scheme significantly improves bandwidth use in complex data center networks.
• It reduces the probability of TCP incast by up to 30% compared to conventional methods like VLB, ECMP, and IFS-RL.

Conclusions:

• The developed TCP flow management scheme effectively optimizes bandwidth use in complex data center networks.
• It offers a superior solution for managing network traffic and mitigating incast issues.