Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Communication-Efficient Training Workload Balancing for Decentralized Multi-Agent Learning

Seyed Mahmoud Sajjadi Mohammadabadi, Lei Yang, Feng Yan, Junshan Zhang

TL;DR

A Communication-Efficient Training Workload Balancing for Decentralized Multi-Agent Learning (ComDML), which balances the workload among agents through a decentralized approach, and demonstrates robustness in heterogeneous environments, and privacy measures can be seamlessly integrated for enhanced data protection.

Abstract

Decentralized Multi-agent Learning (DML) enables collaborative model training while preserving data privacy. However, inherent heterogeneity in agents' resources (computation, communication, and task size) may lead to substantial variations in training time. This heterogeneity creates a bottleneck, lengthening the overall training time due to straggler effects and potentially wasting spare resources of faster agents. To minimize training time in heterogeneous environments, we present a Communication-Efficient Training Workload Balancing for Decentralized Multi-Agent Learning (ComDML), which balances the workload among agents through a decentralized approach. Leveraging local-loss split training, ComDML enables parallel updates, where slower agents offload part of their workload to faster agents. To minimize the overall training time, ComDML optimizes the workload balancing by jointly considering the communication and computation capacities of agents, which hinges upon integer programming. A dynamic decentralized pairing scheduler is developed to efficiently pair agents and determine optimal offloading amounts. We prove that in ComDML, both slower and faster agents' models converge, for convex and non-convex functions. Furthermore, extensive experimental results on popular datasets (CIFAR-10, CIFAR-100, and CINIC-10) and their non-I.I.D. variants, with large models such as ResNet-56 and ResNet-110, demonstrate that ComDML can significantly reduce the overall training time while maintaining model accuracy, compared to state-of-the-art methods. ComDML demonstrates robustness in heterogeneous environments, and privacy measures can be seamlessly integrated for enhanced data protection.

Communication-Efficient Training Workload Balancing for Decentralized Multi-Agent Learning

TL;DR

A Communication-Efficient Training Workload Balancing for Decentralized Multi-Agent Learning (ComDML), which balances the workload among agents through a decentralized approach, and demonstrates robustness in heterogeneous environments, and privacy measures can be seamlessly integrated for enhanced data protection.

Abstract

Decentralized Multi-agent Learning (DML) enables collaborative model training while preserving data privacy. However, inherent heterogeneity in agents' resources (computation, communication, and task size) may lead to substantial variations in training time. This heterogeneity creates a bottleneck, lengthening the overall training time due to straggler effects and potentially wasting spare resources of faster agents. To minimize training time in heterogeneous environments, we present a Communication-Efficient Training Workload Balancing for Decentralized Multi-Agent Learning (ComDML), which balances the workload among agents through a decentralized approach. Leveraging local-loss split training, ComDML enables parallel updates, where slower agents offload part of their workload to faster agents. To minimize the overall training time, ComDML optimizes the workload balancing by jointly considering the communication and computation capacities of agents, which hinges upon integer programming. A dynamic decentralized pairing scheduler is developed to efficiently pair agents and determine optimal offloading amounts. We prove that in ComDML, both slower and faster agents' models converge, for convex and non-convex functions. Furthermore, extensive experimental results on popular datasets (CIFAR-10, CIFAR-100, and CINIC-10) and their non-I.I.D. variants, with large models such as ResNet-56 and ResNet-110, demonstrate that ComDML can significantly reduce the overall training time while maintaining model accuracy, compared to state-of-the-art methods. ComDML demonstrates robustness in heterogeneous environments, and privacy measures can be seamlessly integrated for enhanced data protection.
Paper Structure (25 sections, 2 theorems, 21 equations, 3 figures, 3 tables, 1 algorithm)

This paper contains 25 sections, 2 theorems, 21 equations, 3 figures, 3 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Background and Related Work
  3. Collaborative Multi-agent Learning
  4. Server-less and Peer-to-Peer Decentralized Learning
  5. Workload Balancing for Decentralized Multi-Agent Learning
  6. Decentralized Multi-agent Learning
  7. Workload Balancing via Local-loss based Split Training
  8. Optimization for Workload Balancing
  9. Decentralized Workload Balancing
  10. Agent Pairing
  11. Training Workflow
  12. Privacy Protection
  13. Convergence Analysis
  14. Experimental Evaluation
  15. Experimental Setup
  16. ...and 10 more sections

Key Result

Theorem 1

Suppose that $f^{a_s^m}$ and $f^{a_f^m}$ satisfy Assumptions as:l_smooth, as:bound_gradient, as:bound_variance, as:gradient_diss, and as:conv_pre_layer. The convergence properties of ComDML for both convex and non-convex functions are summarized as follows: where $H_1$, $H_2$, $D$, $F^{a_s^m}$, $F^{a_f^m}$, and $A^{m}$ are constants whose definitions are provided in the Appendix for reference. De

Figures (3)

  • Figure 1: Comparison of model training with and without workload balancing. Workload balancing reduces training time by offloading the workload from agent 1 to agent 2, which would otherwise be idle.
  • Figure 2: Overview of the training process for ComDML. In each round, ComDML first pairs agents ①. The paired agents then carry out the local model update ②. After all agents complete the model update, ComDML uses a decentralized model aggregation to obtain the global model ③. In the above example, agents 1 and 2 offload work to faster agents 5 and 4, respectively, while agent 3 trains independently.
  • Figure 3: Comparison of total training time (in seconds) against baseline models under a 20% link connectivity in random topology.

Theorems & Definitions (2)

  • Theorem 1: Convergence of ComDML
  • Lemma 1