K8S Cpu-Manager-Policy

注:K8S版本1.10
K8S使用過(guò)程中發(fā)現(xiàn)java程序啟動(dòng)特別慢的情況,這個(gè)原因主要是由于K8S CPU上下文的頻繁切換導(dǎo)致的CPU利用率不高?？梢酝ㄟ^(guò)K8S的cpu-manager-policy來(lái)解決這個(gè)問(wèn)題，配置如下。

在kubelet的啟動(dòng)參數(shù)中加入如下參數(shù):

--feature-gates=CPUManager=true --cpu-manager-policy=static --kube-reserved=cpu=1,memory=0

重啟kubelet

systemctl restart kubelet

發(fā)現(xiàn)報(bào)錯(cuò)

Jan 23 10:49:30 sh-saas-k8s1-node-dev-02 systemd: Starting Kubernetes systemd probe.
Jan 23 10:49:30 sh-saas-k8s1-node-dev-02 kubelet: panic: [cpumanager] state file: unable to restore state from disk (policy configured "static" != policy from state file "none")
Jan 23 10:49:30 sh-saas-k8s1-node-dev-02 kubelet: Panicking because we cannot guarantee sane CPU affinity for existing containers.
Jan 23 10:49:30 sh-saas-k8s1-node-dev-02 kubelet: Please drain this node and delete the CPU manager state file "/data/kubernetes/kubelet/cpu_manager_state" before restarting Kubelet.
Jan 23 10:49:30 sh-saas-k8s1-node-dev-02 kubelet: goroutine 1 [running]:

這個(gè)是由于kubelet有個(gè)state文件，按照上面的找到那個(gè)文件，手動(dòng)把狀態(tài)改成static即可

重新啟動(dòng)kubelet，將需要高性能的應(yīng)用的limit和request設(shè)置成一樣，且cpu個(gè)數(shù)必須為整數(shù)個(gè)數(shù)，這樣K8S會(huì)將應(yīng)用自動(dòng)進(jìn)行cpu綁定，性能會(huì)有一個(gè)比較大的提升。