MIG Reconfiguration

MIG Manager draining workloads and repartitioning A100...

GPU Sharing — Live Simulation
8:00 AM
DAY
idle · select workload or Run to simulate
Management Plane Data Plane / User Journey Observability deploys manages ArgoCD GitOps deploy GPU Operator v24.9.2 · CRI-O MIG Manager DaemonSet audio-and-llm 2×1g + 1×2g + 1×3g 10 Clients click to expand Audio 1 · P1 Audio 2 · P1 Audio 3 · P1 LLM-sm 1 · P2 LLM-sm 2 · P2 LLM-lg 1 · P2 LLM-lg 2 · P2 Batch 1 · P3 Batch 2 · P3 Batch 3 · P3 0 active · 0 queued Turnstone agent loop (opt.) OCP Service ClusterIP · typed resource routing nvidia.com/mig-* Triton whisper #1 0/2 Triton whisper #2 0/2 Triton vLLM 8B 0/1 Triton vLLM 22B 0/1 A100-80GB MIG enabled · mixed strategy 1g.10gb 10 GB · whisper 1g.10gb 10 GB · whisper 2g.20gb 20 GB · LLM 8B 3g.40gb 40 GB · LLM 22B SM util: hardware isolated dedicated SMs + HBM2e DCGM Exporter (DaemonSet) requests: 0 active temp: --°C · power: -- W ⚡ P3 preempted by P1 reconfiguring MIG... Datadog Agent DaemonSet · NVML Datadog metrics · monitors + alerts ALERT Pool · llm-small Pool · llm-large Gateway Envoy · model routing Endpoint Picker KV-cache aware skips llm-d Management Data / Journey Observability P1 Audio P2 LLM P3 Batch Click any node to explore
⚙
Component
detail