K8S原理剖析：调度器原理剖析和实践

不念不忘少年蓝@ 2023-06-26 12:20 3阅读 0赞

# 大纲 #

>  *  K8S调度机制介绍
>  *  K8S中的调度策略与算法
>  *  K8S高级调度特性详解

# K8S调度机制介绍 #

## Scheduler：为Pod找到一个合适的Node ##

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ_size_16_color_FFFFFF_t_70][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ_size_16_color_FFFFFF_t_70 1][]

## Kubernetes的Default scheduler ##

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ_size_16_color_FFFFFF_t_70 2][]

## 从外部流程看调度器 ##

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ_size_16_color_FFFFFF_t_70 3][]

## 调度器的内部流程 ##

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ_size_16_color_FFFFFF_t_70 4][]

> 1.  通过NodeLister获取所有节点信息；
> 2.  整合scheduled pods和assume pods，合并到pods，作为所有已调度Pod信息；
> 3.  从pods中整理出node-pods的对应关系表nodeNameToInfo；
> 4.  过滤掉不合适的节点；
> 5.  给剩下的节点依次打分；
> 6.  在分数最高的nodes中随机选择一个节点用于绑定。这是为了避免分数最高的节点被几次调度撞车。

# K8S中的调度策略与算法 #

## K8S中的调度策略与算法     ##

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ_size_16_color_FFFFFF_t_70 5][]

## 通过Predicate策略筛选符合条件的Node ##

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ_size_16_color_FFFFFF_t_70 6][]

### 典型Predicate算法 ###

<table> 
 <tbody> 
  <tr> 
   <td>算法名称</td> 
   <td>功能</td> 
  </tr> 
  <tr> 
   <td>GeneralPredicates</td> 
   <td>包含3项基本检查： 节点、端口和规则</td> 
  </tr> 
  <tr> 
   <td>NoDiskConflict</td> 
   <td>检查Node是否可以满足Pod对硬盘的需求</td> 
  </tr> 
  <tr> 
   <td>NoVolumeZoneConflict</td> 
   <td>单集群跨AZ部署时，检查node所在的zone是否能满足Pod对硬盘的需求</td> 
  </tr> 
  <tr> 
   <td>MaxEBSVolumeCount</td> 
   <td>部署在AWS时，检查node是否挂载了太多EBS卷</td> 
  </tr> 
  <tr> 
   <td>MaxGCEPDVolumeCount</td> 
   <td>部署在GCE时，检查node是否挂载了太多PD卷</td> 
  </tr> 
  <tr> 
   <td>PodToleratesNodeTaints</td> 
   <td>检查Pod是否能够容忍node上所有的taints</td> 
  </tr> 
  <tr> 
   <td>CheckNodeMemoryPressure</td> 
   <td>当Pod QoS为besteffort时，检查node剩余内存量， 排除内存压力过大的node</td> 
  </tr> 
  <tr> 
   <td>MatchInterPodAffinity</td> 
   <td>检查node是否满足pod的亲和性、反亲和性需求</td> 
  </tr> 
 </tbody> 
</table>

## 通过Priority策略给剩余的Node评分，挑选最优的节点 ##

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ_size_16_color_FFFFFF_t_70 7][]

### 典型Priority算法 ###

<table> 
 <tbody> 
  <tr> 
   <td>算法名称</td> 
   <td>功能</td> 
  </tr> 
  <tr> 
   <td>LeastRequestedPriority</td> 
   <td>按node计算资源(CPU/MEM)剩余量排序，挑选最空闲的node</td> 
  </tr> 
  <tr> 
   <td>BalancedResourceAllocation</td> 
   <td>补充LeastRequestedPriority，在cpu和mem的剩余量取平衡</td> 
  </tr> 
  <tr> 
   <td>SelectorSpreadPriority</td> 
   <td>同一个Service/RC下的Pod尽可能的分散在集群中。 Node上运行的同个Service/RC下的<br> Pod数目越少，分数越高。</td> 
  </tr> 
  <tr> 
   <td>NodeAffinityPriority</td> 
   <td>按soft(preferred) NodeAffinity规则匹配情况排序，规则命中越多，分数越高</td> 
  </tr> 
  <tr> 
   <td>TaintTolerationPriority</td> 
   <td>按pod tolerations与node taints的匹配情况排序，越多的taints不匹配，分数越低</td> 
  </tr> 
  <tr> 
   <td>InterPodAffinityPriority</td> 
   <td>按soft(preferred) Pod Affinity/Anti-Affinity规则匹配情况排序，规则命中越多，分数越高<br> /低</td> 
  </tr> 
 </tbody> 
</table>

# K8S高级调度特性详解 #

## K8S中的Label与Selector ##

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ_size_16_color_FFFFFF_t_70 8][]

##  Node Affinity 让Pod在一组指定的Node上运行 ##

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ_size_16_color_FFFFFF_t_70 9][]

## Pod Affinity 让Pod与指定Service的一组Pod在相同Node上运行 ##

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ_size_16_color_FFFFFF_t_70 10][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ_size_16_color_FFFFFF_t_70 11][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ_size_16_color_FFFFFF_t_70 12][]

## Pod Anti-Affinity 让同一个Service的Pod分在到不同Node上运行 ##

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ_size_16_color_FFFFFF_t_70 13][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ_size_16_color_FFFFFF_t_70 14][]

## Pod Anti-Affinity具有对称性 ##

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ_size_16_color_FFFFFF_t_70 15][]

## Taints-tolerations 来自Node的反亲和配置 ##

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ_size_16_color_FFFFFF_t_70 16][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ_size_16_color_FFFFFF_t_70 17][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ_size_16_color_FFFFFF_t_70 18][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ_size_16_color_FFFFFF_t_70 19][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ_size_16_color_FFFFFF_t_70 20][]

[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ_size_16_color_FFFFFF_t_70]: https://img-blog.csdnimg.cn/20200101171016276.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ_size_16_color_FFFFFF_t_70 1]: https://img-blog.csdnimg.cn/20200101171044101.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ_size_16_color_FFFFFF_t_70 2]: https://img-blog.csdnimg.cn/20200101171139256.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ_size_16_color_FFFFFF_t_70 3]: https://img-blog.csdnimg.cn/20200101171833650.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ_size_16_color_FFFFFF_t_70 4]: https://img-blog.csdnimg.cn/20200101171935516.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ_size_16_color_FFFFFF_t_70 5]: https://img-blog.csdnimg.cn/202001011732004.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ_size_16_color_FFFFFF_t_70 6]: https://img-blog.csdnimg.cn/20200101173256938.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ_size_16_color_FFFFFF_t_70 7]: https://img-blog.csdnimg.cn/20200101173357239.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ_size_16_color_FFFFFF_t_70 8]: https://img-blog.csdnimg.cn/2020010117363828.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ_size_16_color_FFFFFF_t_70 9]: https://img-blog.csdnimg.cn/20200101173847645.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ_size_16_color_FFFFFF_t_70 10]: https://img-blog.csdnimg.cn/20200101174503347.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ_size_16_color_FFFFFF_t_70 11]: https://img-blog.csdnimg.cn/20200101180518949.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ_size_16_color_FFFFFF_t_70 12]: https://img-blog.csdnimg.cn/20200101180622190.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ_size_16_color_FFFFFF_t_70 13]: https://img-blog.csdnimg.cn/20200101180640936.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ_size_16_color_FFFFFF_t_70 14]: https://img-blog.csdnimg.cn/20200101180800710.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ_size_16_color_FFFFFF_t_70 15]: https://img-blog.csdnimg.cn/20200101180902138.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ_size_16_color_FFFFFF_t_70 16]: https://img-blog.csdnimg.cn/20200101181009553.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ_size_16_color_FFFFFF_t_70 17]: https://img-blog.csdnimg.cn/20200101181116992.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ_size_16_color_FFFFFF_t_70 18]: https://img-blog.csdnimg.cn/20200101181144196.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ_size_16_color_FFFFFF_t_70 19]: https://img-blog.csdnimg.cn/20200101181328578.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ_size_16_color_FFFFFF_t_70 20]: https://img-blog.csdnimg.cn/20200101181410503.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2ZseTkxMDkwNQ==,size_16,color_FFFFFF,t_70