Dynamic Graph Partitioning for Distributed Deep Learning

Abstract

Distributed deep learning has emerged as a crucial paradigm for training increasingly complex models, often necessitating the partitioning of computational graphs across multiple devices. However, traditional graph partitioning techniques rely on static heuristics, failing to adapt to the dynamic and evolving workload demands inherent in distributed training. This paper introduces a novel approach utilizing reinforcement learning (RL) to dynamically partition the graph, addressing this limitation. The proposed system employs an RL agent that observes the current state of the training process – including communication costs and computational load – and dynamically adjusts the graph partitioning scheme. This adaptive partitioning optimizes for efficiency, leading to improved training performance and reduced communication overhead. The core of the method lies in the agent's ability to learn optimal partitioning strategies through trial and error, creating a system that responds effectively to fluctuations in the training workload. The system is evaluated conceptually, outlining key components and expected benefits. Further research and experimentation are planned to fully validate these concepts.