Conectar

Global Capacity Orchestrator on AWS

aws-solutions-library-samplesbyaws-solutions-library-samples
•
53
•
Ciencia de Datos y ML
Infraestructura en la Nube
Despliegue y DevOps

Orchestrate distributed AI/ML and HPC workloads across multi-region AWS infrastructure with capacity-aware scheduling and unified management.

Global Capacity Orchestrator (GCO) simplifies running large-scale accelerated workloads—such as LLM training, distributed inference, batch machine learning, and high-performance computing—across multiple AWS Regions using Amazon EKS Auto Mode, AWS PCS (Slurm), and SageMaker HyperPod. It features automated discovery for NVIDIA GPUs, Trainium, and Inferentia capacity, enabling optimal job placement and automated cross-region inference failover via AWS Global Accelerator. With unified AWS IAM authentication, native CLI support, and a built-in Model Context Protocol (MCP) server, it allows AI agents and engineering teams to manage global computational capacity without the friction of per-cluster credential management.

Características Principales

01Unified IAM-authenticated API, CLI, and MCP server eliminating kubeconfig distribution
02On-demand multi-accelerator provisioning covering NVIDIA GPUs, Trainium, Inferentia, and EFA
03Capacity-aware job placement analyzing spot placement scores, reservations, and training plans
0453 GitHub stars
05Pre-integrated AI and HPC schedulers including Volcano, Kueue, KubeRay, Kubeflow Trainer, and Slurm
06Multi-region LLM and model inference deployments with automatic failover via Global Accelerator

Casos de Uso

01Locating and scheduling large-scale GPU or Trainium model training jobs where capacity is readily available
02Deploying resilient, multi-region LLM inference endpoints with automatic latency routing and failover
03Enabling AI coding agents like Claude Code to orchestrate infrastructure operations via the integrated MCP server