System and Method for Generative AI-Driven Multi-Layer Topology Awareness and Dynamic Workload Placement in Container Orchestration Platforms

Abstract

Modern container orchestration platforms schedule workloads with limited awareness of the physical and logical topology underlying the cluster — rack placement, network fabric hierarchy, NUMA and PCIe locality, GPU interconnect (NVLink/NVSwitch) topology, power distribution, thermal state, and multi-cloud or multi-region link characteristics. Existing solutions such as Node Feature Discovery and the native Kubernetes Topology Manager address this problem with static, local, rule-based heuristics that are evaluated on a single node at a time, and they treat placement as a one-time decision made at scheduling time with no subsequent runtime feedback. This disclosure describes a system and method that builds a multi-layer topology graph spanning physical, network, accelerator-interconnect, power, and thermal layers, encodes it into continuous spatio-temporal graph embeddings, and uses a generative AI engine to evaluate counterfactual placement states and produce an optimized Placement Policy Graph for each scheduling request. A Closed-Loop Performance Verification and Adaptation Engine then continuously compares actual post-placement execution metrics against the model's predictions; when the deviation exceeds a configured threshold, the system automatically executes real-time micro-adjustments — including updating eBPF network routing priorities, modifying accelerator power caps via NVML, or triggering live sub-task migrations — and feeds the observed deviation back into the generative model for continued fine-tuning. We describe the system architecture, the graph representation and candidate-generation algorithm underlying it, and a concrete evaluation methodology — including real baselines, public cluster traces, and ablations — for validating the approach empirically. The result is a predictive, generative, and self-healing placement engine that adapts to heterogeneous, GPU-dense, multi-layer environments without hand-tuned scheduling heuristics for each new topology shape, and that continues correcting itself after a workload is already running.