kubean-io/kube-node-tuning

Manage kubernetes node-level kernel tuning ( using sysctl ).

kube-node-tuning

kube-node-tuning is aimed at kernel tuning by kubernetes. It's useful to :

• High performance application
• Large Scale Cluster
• Network Tuning

Quick Start

export VERSION=v0.3.1
helm repo add kube-node-tuning https://kubean-io.github.io/kube-node-tuning/
helm install -n kube-node-tuning kube-node-tuning kube-node-tuning/kube-node-tuning --version $VERSION --create-namespace

Tips: If the machine is in China, you should following by: quick-start-in-china

The kernel's sysctl settings are applied to the node at /etc/99-kube-node-tuning.conf.

Check if the settings are applied by the following command. SSH to the node of cluster

cat /etc/sysctl.d/99-kube-node-tuning.conf
sysctl -a # view the sysctl setting

Configuration

# Change the config
kubectl -n kube-node-tuning edit cm/kube-node-tuning-config -o yaml

# Restart the DaemonSet
kubectl -n kube-node-tuning rollout restart ds kube-node-tuning

Roadmap

• Different OS Support. (Ubuntu, CentOS, RHEL, etc.)
• Multi profile
• Operator instead of Daemonset

kube-node-tuning's Issues

First Release

梳理 sysctl 参数

优化自带的参数，并完善文档

Helm Support

optimize arp parameters

when stress testing, the network presure increase.This may cause arp overflow.
The following parameters need to be considered：

net.ipv4.neigh.default.gc_thresh1 = 8192
net.ipv4.neigh.default.gc_thresh2 = 32768
net.ipv4.neigh.default.gc_thresh3 = 65536

Support max_user_instances

fs.inotify.max_user_watches=65536
fs.inotify.max_user_instances=8192

Like:

https://github.com/openshift/cluster-node-tuning-operator/blob/master/assets/tuned/daemon/profiles/openshift-node/tuned.conf

add CI test

Init Design

Kernel Tunning is very useful in Production Kubernetes. Like:

Network:

• kubernetes-sigs/kubespray#8825

And OpenShift use another project to solve the problem. But it can only run in RHEL/Centos.

https://github.com/openshift/cluster-node-tuning-operator

We need a solution, which can run any on any OS to config the sysctl.

Refer:

• https://dwiki.daocloud.io/pages/viewpage.action?pageId=45453501
• https://dwiki.daocloud.io/pages/viewpage.action?pageId=27500572

Use Github CI to release

Uninstall yaml

Use initContainer instead of nomarl container

Use Busybox base image

将脚本 打包成镜像

是否支持更改 configmap 之后热重启自动生效

package by it's own image not busybox.

add e2e test

Multi profile for multi use-case

job模式，这样就没有常驻的 POD了。

daemonset can not start

pid_max optimization

when /proc/sys/kernel/pid_max is small (such as 65536) , it may give rise to the upper application report "Cannot allocate memory".

撰写 readme

更完善一些

add dependency bot