k8s memory manager deep dive

发表于 更新于 本文字数： 774 阅读时长 ≈ 1 分钟

k8s-memory-manager-deep-dive

  • 只能针对GU, 不保证burstable的pod是否跨numa访问.
  • kube-scheduler无法感知numa: The default Kubernetes scheduler is not aware of the node’s NUMA topology, and it can be a reason for many admission errors during the pod start.

疑问汇总

  1. mm怎么跟cadvisor交互, 来获取内存信息? 详细代码实现. ❌

  2. 如下图, 如果到操作系统来给pod发送OOM, 如何保证 “Low-priority containers are targeted before the high-priority containers”? 在进程优先级上有设定? 底层是怎么实现的? ✅

答案: 参见 Node-pressure Eviction | Kubernetes 底层是给不同QoS的Pod设定不同的oom_score_adj, 然后通过Linux的 oom_killer 识别&执行.

Untitled

  1. 为啥下边的场景, Pod2创建会失败? ❌

Untitled

  1. 怎么跟cpumanager进行交互?

Refs