Holistic Load Balancing for Heterogeneous AI Infrastructures: A Graph Neural Network Approach Article Swipe

MAKE IMPACT

Revista, Zen , IA, 10 · YOU? · · 2025 · Open Access · · DOI: https://doi.org/10.5281/zenodo.17822484

The proliferation of Artificial Intelligence (AI) workloads, ranging from large-scale model training to real-time inference, presents unprecedented challenges for underlying computational infrastructures. These infrastructures are increasingly heterogeneous, comprising diverse accelerators such as GPUs, FPGAs, TPUs, and specialized AI ASICs, each with distinct capabilities and optimal performance profiles. Traditional load balancing strategies, often designed for homogeneous general-purpose computing, fall short in effectively managing the complex interplay of diverse AI tasks and varied hardware resources. This paper proposes a novel, holistic load balancing approach for heterogeneous AI infrastructures leveraging Graph Neural Networks (GNNs). We model the entire infrastructure—including compute nodes, network topology, pending AI tasks, and task dependencies—as a dynamic graph. The GNN is trained to learn complex spatial and temporal relationships within this graph, enabling it to make informed, adaptive resource allocation and task scheduling decisions that optimize for multiple objectives such as minimizing task completion time, maximizing resource utilization, and enhancing energy efficiency. Our methodology details the GNN architecture, feature engineering for heterogeneous environments, training strategies using simulated and real-world data, and comprehensive evaluation metrics. Initial findings, derived from extensive simulations, demonstrate that the GNN-based approach significantly outperforms conventional and heuristic-based load balancing methods by achieving superior resource utilization, reduced task latency, and improved overall system throughput in dynamic, heterogeneous AI environments. This research contributes a robust and scalable framework for intelligent resource management in the rapidly evolving landscape of AI computing.

Type

article

Landing Page

https://doi.org/10.5281/zenodo.17822484

OA Status

green

OpenAlex ID

https://openalex.org/W7108731992

Raw OpenAlex JSON

OpenAlex ID

https://openalex.org/W7108731992

Canonical identifier for this work in OpenAlex

DOI

https://doi.org/10.5281/zenodo.17822484

Digital Object Identifier

Title

Holistic Load Balancing for Heterogeneous AI Infrastructures: A Graph Neural Network Approach

Work title

Type

article

OpenAlex work type

Publication year

2025

Year of publication

Publication date

2025-12-04

Full publication date if available

Authors

Revista, Zen, IA, 10

List of authors in order

Landing page

https://doi.org/10.5281/zenodo.17822484

Publisher landing page

Open access

Yes

Whether a free full text is available

OA status

green

Open access status per OpenAlex

OA URL

https://doi.org/10.5281/zenodo.17822484

Direct OA link when available

Concepts

Computer science, Load balancing (electrical power), Scalability, Distributed computing, Scheduling (production processes), Artificial neural network, Artificial intelligence, Task (project management), Machine learning, Graph, Robustness (evolution), Homogeneous, Heterogeneous network, Adaptation (eye), Task analysis, Resource (disambiguation), Feature (linguistics), Job shop scheduling, Resource allocation, Load management, Resource efficiency, Feature engineering, Resource management (computing)

Top concepts (fields/topics) attached by OpenAlex

Cited by

0

Total citation count in OpenAlex

Full payload