Dual Plane Orchestration Framework Enhances Distributed AI Pipeline Management Across Networks

1 min read
A new framework separates AI workloads into inference-focused data plane and parameter-focused control plane, enabling efficient orchestration of distributed AI functions across heterogeneous edge-cloud infrastructures.

Researchers led by Roberto Riggio have developed a Dual Plane Orchestration Framework to optimize deployment of distributed artificial intelligence (AI) pipelines across converged connect-compute networks. Published in the Journal of Network and Systems Management, the framework introduces Artificial Intelligence Functions as modular components representing individual machine learning pipeline stages.

The innovation lies in explicitly separating AI workloads into two planes: a latency-sensitive data plane that handles continuous inference traffic, and a latency-tolerant control plane that manages bulk model parameter updates. This distinction addresses conflicting communication requirements inherent in distributed AI workflows spanning edge devices to centralized clouds.

The framework models application requirements as directed graphs and maps them onto physical substrate networks with multi-dimensional resource capacities. An integer linear programming formulation minimizes placement costs under capacity and latency constraints. To enable scalability, a two-phase greedy heuristic algorithm was developed, achieving near-optimal resource placement with execution times three to four orders of magnitude faster than exact solvers.

Evaluations across varied network sizes and request complexities demonstrated that this heuristic outperforms existing orchestration techniques by reducing embedding costs and increasing acceptance ratios. Explicit differentiation between data and control planes accounted for significant performance gains. The approach proved scalable and practical for next-generation edge-cloud AI deployments.

An open-source, cloud-native software prototype integrating with Kubernetes and the LightEdge platform confirmed the framework’s deployment feasibility, achieving orchestration latencies under 350 milliseconds in test scenarios.

This work provides a computationally efficient foundation for orchestrating complex multi-stage AI workflows across heterogeneous infrastructures, with implications for sectors including automated mobility, healthcare, and industrial automation.

Source: Bioengineer.org.

Source