Core Concepts
Leveraging sparse mean field theory to develop near-optimal load balancing policies in large queueing systems.
Abstract
The article discusses the challenges of modeling large queueing networks with strong locality and proposes a novel approach using sparse mean field theory. It addresses the need for scalable load balancing algorithms in cloud networks and data centers, focusing on reducing job drops in sparsely connected queueing systems. The content is structured into sections covering the introduction, system model, locality and scalability, sparse graph mean field system, reinforcement learning, training on a finite queueing system, and optimality guarantees.
Introduction:
Scalable load balancing algorithms are crucial for cloud networks and data centers.
Existing techniques struggle to model large queueing networks with strong locality.
Proposal to leverage sparse mean field theory for near-optimal load balancing policies.
System Model:
Describes a decentralized system with agents and servers.
Jobs processed at exponential rates with finite buffer capacity.
Agents access limited queues based on predefined topologies.
Locality and Scalability:
Focuses on sparsely connected queueing systems with limited agent access to queues.
Strong concept of locality implies partial observability in the system.
Sparse Graph Mean Field System:
Introduces technical details for convergence in the local weak sense.
Defines limiting topology of large systems using bounded-degree graphs.
Reinforcement Learning:
Formulates MFC problem as an MDP for single-agent RL.
Considers partially observed MDP variant due to complexity of infinite-sized systems.
Training on a Finite Queueing System:
Discusses training POMDP using proximal policy optimization (PPO) RL method.
Centralized training decentralized execution scheme applied to localized queueing systems.
Optimality Guarantees:
Theoretical guarantees show performance convergence between finite and limiting systems.
Stats
"Scalable load balancing algorithms are of great interest in cloud networks and data centers."
"Empirically, the proposed methodology performs well on several realistic and scalable wireless network topologies."
Quotes
"In this work, we address this challenge by leveraging recent advances in sparse mean field theory."
"Our main contributions through this work are..."