GKE
Google Cloud's managed Kubernetes service, used to deploy and manage containerized applications.
Timeline (last 90 days)
GKE's GPU metrics come from Google's in-house nvidia-metrics-collector container, which runs inside the device-plugin DaemonSet rather than the community-standard dcgm-exporter.
GKE managed GPU metrics export only four metrics: duty_cycle, memory_used, memory_total, and request.
GKE managed GPU metrics do not include SM occupancy, Tensor Core activity, or power consumption data.
GKE's duty_cycle samples carry the pod, namespace, and container information of the actual workload, enabling accurate attribution to specific business units.
The GPU metrics exported by GKE come from Google's in-house nvidia-metrics-collector container, rather than the community-standard dcgm-exporter.
The GKE managed solution has correct attribution but misleading and non-upgradeable utilization metrics, while self-built DCGM has incorrect attribution by default but honest and configurable utilization metrics.
The GKE managed GPU collector exports only four metrics: duty_cycle, memory_used, memory_total, and request, with no SM occupancy, Tensor Core activity, or power consumption data.
It is not possible to enable any finer-grained metrics on GKE managed metrics; kernel residency is the only utilization signal exported.
Each duty_cycle sample from the GKE managed collector carries the pod, namespace, and container information of the actual workload, making attribution accurate.
All Facts (10)
谷歌云以BigQuery（数据分析）、Vertex AI（AI/ML）和GKE（Kubernetes容器编排）作为企业级市场的核心技术卖点
85%UnverifiedGKE's duty_cycle samples carry the pod, namespace, and container information of the actual workload, enabling accurate attribution to specific business units.
50%UnverifiedGKE's GPU metrics come from Google's in-house nvidia-metrics-collector container, which runs inside the device-plugin DaemonSet rather than the community-standard dcgm-exporter.
50%UnverifiedGKE managed GPU metrics export only four metrics: duty_cycle, memory_used, memory_total, and request.
50%UnverifiedGKE managed GPU metrics do not include SM occupancy, Tensor Core activity, or power consumption data.
50%UnverifiedThe GKE managed solution has correct attribution but misleading and non-upgradeable utilization metrics, while self-built DCGM has incorrect attribution by default but honest and configurable utilization metrics.
50%UnverifiedThe GPU metrics exported by GKE come from Google's in-house nvidia-metrics-collector container, rather than the community-standard dcgm-exporter.
50%UnverifiedThe GKE managed GPU collector exports only four metrics: duty_cycle, memory_used, memory_total, and request, with no SM occupancy, Tensor Core activity, or power consumption data.
50%UnverifiedIt is not possible to enable any finer-grained metrics on GKE managed metrics; kernel residency is the only utilization signal exported.
50%UnverifiedEach duty_cycle sample from the GKE managed collector carries the pod, namespace, and container information of the actual workload, making attribution accurate.
50%