Monitoring

A complete monitoring dashboard includes high-level and low-level views of the cluster, providing comprehensive visibility into system performance, resource utilization, and potential issues before they impact workloads.

Cluster Overview

  • Number of nodes, pods, workloads running
  • Current node health status
  • Resource utilization (CPU, memory, disk, network) at cluster/node/namespace/pod levels
  • Trends and summaries
  • Workload stability (restarts, pending pods, scaling events)
  • Control plane health (API server latency, etcd performance, controller manager and scheduler metrics)

Resource Management Integration

  • Slurm integration with job stats by user, type, etc.Slurm Accounting
  • Kubernetes integration with kube-state-metrics, node-exporter, dcgm-exporter, cAdvisordcgm-exporter
  • Resource quotas with limits and usage of GPUs for users and groups
  • Real-time GPU node availability status
  • Scheduler resource contention monitoring

Monitoring Stack

  • kube-prometheus-stack for Kubernetes metrics (node state, resource usage, control plane, workload performance)kube-prometheus-stack
  • NVIDIA DCGM integration for comprehensive GPU monitoring or AMD equivalentDCGM docs
  • KV cache usage for horizontal pod autoscaling with serving (gpu_cache_usage_perc)
  • Integrated alert management and notification systems

Hardware Monitoring

  • Node-level power draw with real-time power consumption monitoring
  • Fan speed monitoring for cooling system performance and status
  • Temperature monitoring (CPU, RAM, NIC, transceiver, other critical components)
  • PCIe AER rates for Advanced Error Reporting on PCIe bus health
  • dmesg logs monitoring via Promtail for system-level message capturePromtail docs
  • GB200 NVL72 specialized monitoring requirements

Performance Monitoring

  • ncu (Nsight Compute) on PATH so non-root users can profile GPU kernels. Binary check only — counter access is the next item.Nsight Compute CLI
  • GPU hardware counters accessible without sudo: NVreg_RestrictProfilingToAdminUsers=0 set and the driver reloaded, otherwise ERR_NVGPUCTRPERM at runtime even with ncu on PATH. Confirm with a live profile of a trivial kernel, not just the binary.Fixing ERR_NVGPUCTRPERM
  • Setup reference and worked examples. Containers also need SYS_ADMIN (or --privileged) for hardware counters. A full GPU rental should include full profiling access, bare metal, VM, or container.Nsight Compute profiling setup
  • GPU MODE Lecture 44: profiling GPU kernels with Nsight.GPU MODE Lecture 44: NVIDIA Profiling
  • perf top / perf stat usable for CPU-side profiling without root: perf_event_paranoid <= 1 and kptr_restrict = 0 for system-wide and per-PID access. Catches host-side bottlenecks in data loading, preprocessing, and kernel launch overhead.perf_event_paranoid / perf security

Network Monitoring

  • NVLink/XGMI throughput for inter-GPU communication bandwidth
  • PCIe throughput for host CPU to GPU data transfer rates
  • InfiniBand/RoCEv2 throughput for high-speed network performance between nodesRDMA perftest

Resource Management

  • User and group quotas
  • Node availability
  • Scheduler job history

General Expectations

  • Console, dashboard, CLI and/or API available to manage resources
  • 24x7 support availability
  • Process for security fixes and upgrades exists, proactive notifications are clear
  • Integration with active and passive health check systems

All expectations