Running Libri GmbH Scaling AI Systems- Running the GPU Fleet

Keep Your GPU Fleet Fast, Healthy, and Fully UtilizedRunning a modern GPU cluster is a high-stakes challenge. With hardware costs soaring and AI workloads demanding absolute efficiency, cluster operators cannot afford idle time, silent data corruption, or unexpected node failures. Running the GPU Fleet is your definitive, hands-on playbook for operating scale-out AI infrastructure with confidence.Written by experienced cluster engineers, this guide skips the high-level theory to deliver deep, operational blueprints for managing high-performance computing clusters in 2026 and beyond. You will learn how to orchestrate, observe, and maintain accelerators across their entire lifecycle.Inside this comprehensive guide, you will master: - Advanced Kubernetes Orchestration: Configure the NVIDIA GPU Operator, device plugins, MIG profiles, and topology managers without sacrificing predictability.- Efficient Workload Scheduling: Implement gang schedulers like Volcano and Kueue to guarantee atomic, fair-share placements for distributed training.- Deep Observability: Build a robust monitoring stack with DCGM, Prometheus, Grafana, and detailed NCCL profiling to catch errors in minutes.- Failure Management: Troubleshoot frustrating NCCL hangs, InfiniBand link failures, and transceiver degradation using battle-tested playbooks.- Hardware Maintenance & Automation: Detect, isolate, quarantine, and replace faulty nodes across a fleet of thousands of GPUs.Whether you are running Hopper, Blackwell, or next-generation architectures, this book equips you with the exact tools, architectures, and math needed to design multi-tenant clusters that scale seamlessly. Stop wasting compute cycles. Take control of your GPU fleet today!