kubelet最佳配置

2019独角兽企业重金招聘Python工程师标准>>>

author: garnett.wang@gmail.com

kubernetes version: 1.6.2

Kubelet Configurations We Should Care About

下面是我梳理的，我认为必须关注的flag。

flag	value
--address	0.0.0.0
--allow-privileged	false
--cadvisor-port int32	4194
--cgroup-driver string	cgroupfs
--cluster-dns stringSlice	10.0.0.10 //todo
--cluster-domain string	caas.vivo.com
--cni-bin-dir string	/opt/cni/bin
--cni-conf-dir string	/etc/cni/net.d
--docker-endpoint string	unix:///var/run/docker.sock
--eviction-hard string	memory.available<4Gi,<br/> nodefs.available<20Gi,<br/> imagefs.available<5Gi
--eviction-max-pod-grace-period int32	30
--eviction-minimum-reclaim string	memory.available=500Mi,<br/> nodefs.available=2Gi,,<br/> imagefs.available=2Gi
--eviction-pressure-transition-period duration	5m0s
--eviction-soft string	memory.available<8Gi,<br/> nodefs.available<100Gi,<br/> imagefs.available<20Gi
--eviction-soft-grace-period string	memory.available=30s,<br/> nodefs.available=2m,<br/> imagefs.available=2m
--experimental-fail-swap-on	+
--experimental-kernel-memcg-notification	+
--feature-gates string	AllAlpha=false
--file-check-frequency duration	20s
--hairpin-mode string	promiscuous-bridge
--healthz-port int32	10248
--image-gc-high-threshold int32	60
--image-gc-low-threshold int32	40
--image-pull-progress-deadline duration	2m0s
--kube-api-qps int32	5
--kube-reserved mapStringString	cpu=200m,memory=16G
--kubeconfig string	/var/lib/kubelet/kubeconfig
--max-pods int32	50
--minimum-image-ttl-duration duration	1h
--network-plugin string	cni
--pod-infra-container-image string	vivo.registry.com/google_containers/pause-amd64:3.0
--pod-manifest-path string	/var/lib/kubelet/pod_manifest
--port int32	10250
--protect-kernel-defaults	+
--read-only-port int32	10255
--require-kubeconfig	+
--root-dir string	/var/lib/kubelet
--runtime-request-timeout duration	2m0s
--serialize-image-pulls	false
--sync-frequency duration	1m0s
--system-reserved mapStringString	cpu=100m,memory=32G
--volume-plugin-dir string	/usr/libexec/kubernetes/kubelet-plugins/volume/exec/
--volume-stats-agg-period duration	1m0s

Kubelet Configuration Best Practicies

下面是我最终梳理的，认为需要真正显示设置的flag，如下：

/usr/bin/kubelet —address=0.0.0.0 
--port=10250 
--allow-privileged=false 
--cluster-dns=10.0.0.1 
--cluster-domain=caas.vivo.com
--max-pods=50 
--network-plugin=cni 
--require-kubeconfig 
--pod-manifest-path=/etc/kubelet.d/
--pod-infra-container-image=vivo.registry.com/google_containers/pause-amd64:3.0 
--eviction-hard=memory.available<4Gi,nodefs.available<20Gi,imagefs.available<5Gi 
--eviction-max-pod-grace-period=30 
--eviction-minimum-reclaim=memory.available=500Mi,nodefs.available=2Gi,imagefs.available=2Gi 
--eviction-pressure-transition-period=5m0s 
--eviction-soft=memory.available<8Gi,nodefs.available<100Gi,imagefs.available<20Gi 
--eviction-soft-grace-period=memory.available=30s,nodefs.available=2m,imagefs.available=2m 
--experimental-kernel-memcg-notification 
--experimental-fail-swap-on 
--system-reserved=cpu=100m,memory=8G 
--kube-reserved=cpu=200m,memory=16G
--hairpin-mode=promiscuous-bridge 
--image-gc-high-threshold=60 
--image-gc-low-threshold=40 
--serialize-image-pulls=false 
--protect-kernel-defaults 
--feature-gates=AllAlpha=false 

注意：要做好资源预留的话，还需要事先创建以下cgroup目录：

/sys/fs/cgroup/cpuset/system.slice/kubelet.service  
/sys/fs/cgroup/hugebtl/system.slice/kubelet.service  

这个工作可以放到kubelet.service的ExecStartPre中做。

而且一定要加上这三个配置项，否则无法生效：

--enforce-node-allocatable=pods,kube-reserved,system-reserved   
--kube-reserved-cgroup=/system.slice/kubelet.service   
--system-reserved-cgroup=/system.slice  

转载于:https://my.oschina.net/jxcdwangtao/blog/891960