Unverified50% confidenceFactExact time
Kubernetes exposes GPUs as schedulable resources through the Device Plugin mechanism; once a Pod requests a GPU, it exclusively holds that resource until the Pod is destroyed.
1
Sources
50%
Confidence
Long-term
Relevance
8/7/2026
First Seen
Sources
GPU利用率不足30%：预留资源闲置的破局之道
Related Claims
UnverifiedWhen a user launches a Jupyter Notebook and requests GPU resources, the underlying Kernel process holds the GPU memory allocation, and even if the browser tab is closed, the Kernel continues running in the background and locks the GPU.66% similarUnverified当容器内存超过 Pod 资源限制时，Linux 内核的 OOM Killer 会向容器主进程发送不可捕获的 SIGKILL 信号，Kubernetes 将 Pod 标记为 OOMKilled 状态，退出码为 13760% similarUnverifiedMPS不提供故障隔离，单个进程崩溃可能通过MPS daemon传播，拖垮同一GPU上其他任务59% similarUnverifiedKubernetes RBAC 只控制谁能通过 API 创建/读取 Secret 对象，但无法限制 Pod 内部进程对已注入 Secret 的访问57% similarUnverifiedNVIDIA provides the NVIDIA Container Toolkit, enabling Docker containers to directly access the host machine's GPU resources.55% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/702390API
curl https://kongchang.com/api/v1/knowledge/claims/702390MCP
get_claim(id=702390)