[4P2-OS-17b] OS-17

[4P2-OS-17b-01] Microservice Placement Algorithm using Heterogeneous Graph Neural Networks for Supervised Learning

〇Hirotaka Kasahara¹, Ryuichi Kitajima¹, Jun Towada¹, Osamu Sato¹ (1. SoftBank Corp.)

Keywords:Edge Computing, Graph Neural Network, Combinatorial Optimization

With the rise of edge computing and microservice architecture, placing microservices across geographically distributed computing resources while guaranteeing quality of service has become crucial. Existing microservice placement algorithms often fail to adequately consider the complex dependencies within applications and the heterogeneous computing resources in infrastructures, resulting in suboptimal microservice placements. This paper tackles this issue by introducing a novel microservice placement algorithm utilizing supervised learning with heterogeneous graph neural networks. Our approach models both the application structure and infrastructure topology as a unified heterogeneous graph, facilitating effective feature extraction and learning. We utilize relational graph convolutional networks for message aggregation with a modified Dir-GNN and GraphSAGE. Numerical simulation results reveal that our method surpasses existing techniques, achieving a 94.8% accuracy in predicting optimal placements and demonstrating superior computational efficiency. Our method achieves a 1.6% optimality gap in just ten milliseconds calculation, compared to a 7 - 70% gap by metaheuristic methods in 200-500 milliseconds, underscoring its effectiveness and scalability for real-world microservice placement scenarios.