作者 |?leadersnowy

來源 | CSDN博客

kube-scheduler介紹

首先我們知道，kube-scheduler的根本工作任務(wù)是根據(jù)各種調(diào)度算法將Pod調(diào)度到最合適的工作節(jié)點(diǎn)上

一、整個調(diào)度流程分為兩個階段：

1、預(yù)選（Predicates）：

輸入是所有節(jié)點(diǎn)，輸出是滿足預(yù)選條件的節(jié)點(diǎn)。kube-scheduler根據(jù)預(yù)選策略過濾掉不滿足策略的Nodes。例如，如果某節(jié)點(diǎn)的資源不足或者不滿足預(yù)選策略的條件如“Node的label必須與Pod的Selector一致”時則無法通過預(yù)選。

2、優(yōu)選（Priorities）：

輸入是預(yù)選階段篩選出的節(jié)點(diǎn)，優(yōu)選會根據(jù)優(yōu)先策略為通過預(yù)選的Nodes進(jìn)行打分排名，選擇得分最高的Node。例如，資源越富裕、負(fù)載越小的Node可能具有越高的排名。

而整個調(diào)度過程是有很多因素影響的，包括節(jié)點(diǎn)狀態(tài)(cpu，內(nèi)存，磁盤)，節(jié)點(diǎn)與pod的親和性，節(jié)點(diǎn)標(biāo)簽等，而我們本次只做與磁盤相關(guān)的源碼分析。

源碼分析

2.1、磁盤預(yù)選（Predicates）

這一部分核心代碼在pkg/scheduler/framework/plugins/nodevolumelimits/這個目錄下，分兩個文件

csi.go，non_csi.go 兩個文件的核心思想基本一致，都是通過Filter來限制一個節(jié)點(diǎn)上的最大盤的數(shù)量，這個最大數(shù)量是默認(rèn)配置在k8s源碼中的。csi的就是限制一個節(jié)點(diǎn)上最多能掛幾塊csi的volume，而non_csi的相對比較復(fù)雜，因?yàn)闋砍兜絠n-tree插件跟flexvolume插件，而且對每一種in-tree插件都進(jìn)行了分別的可掛載盤的數(shù)量的配置，下面我們對于non-csi的進(jìn)行具體的源碼分析。

話不多說，直接上代碼：

2.1.1 nonCSILimits

type nonCSILimits struct {name stringfilter VolumeFiltervolumeLimitKey v1.ResourceNamemaxVolumeFunc func(node *v1.Node) intcsiNodeLister storagelisters.CSINodeListerpvLister corelisters.PersistentVolumeListerpvcLister corelisters.PersistentVolumeClaimListerscLister storagelisters.StorageClassLister// The string below is generated randomly during the struct's initialization.// It is used to prefix volumeID generated inside the predicate() method to// avoid conflicts with any real volume.randomVolumeIDPrefix string }

最核心的結(jié)構(gòu)體nonCSILimits ，提供了幾個成員

name 顧名思義，每一個nonCSILimits結(jié)構(gòu)體變量的名稱

filter VolumeFilter類型的變量，具體包括幾個方法，FilterVolume(),FilterPersistentVolume(),MatchProvisioner()以及IsMigrated()具體調(diào)用我們后面再講

volumeLimitKey 其實(shí)就是一個string類型，指定了幾種類型的key

maxVolumeFunc() 一個獲取該Limits的最大數(shù)量的方法

csiNodeLister csinode的監(jiān)聽對象

pvLister pv的監(jiān)聽對象

pvcLister pvc的監(jiān)聽對象

scLister sc的監(jiān)聽對象

randomVolumeIDPrefix 一個string類型，用于生成唯一的pvID

2.1.2 nonCSILimits初始化

func newNonCSILimits(filterName string,csiNodeLister storagelisters.CSINodeLister,scLister storagelisters.StorageClassLister,pvLister corelisters.PersistentVolumeLister,pvcLister corelisters.PersistentVolumeClaimLister, ) framework.Plugin {var filter VolumeFiltervar volumeLimitKey v1.ResourceNamevar name stringswitch filterName {case ebsVolumeFilterType:name = EBSNamefilter = ebsVolumeFiltervolumeLimitKey = v1.ResourceName(volumeutil.EBSVolumeLimitKey)case gcePDVolumeFilterType:name = GCEPDNamefilter = gcePDVolumeFiltervolumeLimitKey = v1.ResourceName(volumeutil.GCEVolumeLimitKey)case azureDiskVolumeFilterType:name = AzureDiskNamefilter = azureDiskVolumeFiltervolumeLimitKey = v1.ResourceName(volumeutil.AzureVolumeLimitKey)case indiskVolumeFilterType:name = IndiskNamefilter = indiskVolumeFiltervolumeLimitKey = v1.ResourceName(volumeutil.IndiskVolumeLimitKey)default:klog.Fatalf("Wrong filterName, Only Support %v %v %v %v %v", ebsVolumeFilterType,gcePDVolumeFilterType, azureDiskVolumeFilterType, cinderVolumeFilterType, indiskVolumeFilterType)return nil}pl := &nonCSILimits{name: name,filter: filter,volumeLimitKey: volumeLimitKey,maxVolumeFunc: getMaxVolumeFunc(filterName),csiNodeLister: csiNodeLister,pvLister: pvLister,pvcLister: pvcLister,scLister: scLister,randomVolumeIDPrefix: rand.String(32),}return pl }

初始化時引用了大量的常量

const (// defaultMaxGCEPDVolumes defines the maximum number of PD Volumes for GCE.// GCE instances can have up to 16 PD volumes attached.defaultMaxGCEPDVolumes = 16// defaultMaxAzureDiskVolumes defines the maximum number of PD Volumes for Azure.// Larger Azure VMs can actually have much more disks attached.// TODO We should determine the max based on VM sizedefaultMaxAzureDiskVolumes = 16// ebsVolumeFilterType defines the filter name for ebsVolumeFilter.ebsVolumeFilterType = "EBS"// gcePDVolumeFilterType defines the filter name for gcePDVolumeFilter.gcePDVolumeFilterType = "GCE"// azureDiskVolumeFilterType defines the filter name for azureDiskVolumeFilter.azureDiskVolumeFilterType = "AzureDisk"// cinderVolumeFilterType defines the filter name for cinderVolumeFilter.indiskVolumeFilterType = "Indisk"// ErrReasonMaxVolumeCountExceeded is used for MaxVolumeCount predicate error.ErrReasonMaxVolumeCountExceeded = "node(s) exceed max volume count"// KubeMaxPDVols defines the maximum number of PD Volumes per kubelet.KubeMaxPDVols = "KUBE_MAX_PD_VOLS"// MaxIndiskVolumes defines the maximum number of Indisk Volumes per node.MaxIndiskVolumes = "MAX_INDISK_VOLUMES"// DefaultIndiskVolumes defines the default number of Indisk Volumes per node.DefaultIndiskVolumes = "DEFAULT_INDISK_VOLUMES" )

這些常量包括各intree存儲類型的名字，以及默認(rèn)的最大volume數(shù)量等信息，在nonCSILimits初始化時進(jìn)行賦值。

除了這些常量之外，最關(guān)鍵的就是filter方法的初始化,看一下cinder的例子，是怎么區(qū)分這個volume是cinder的volume的。

var cinderVolumeFilter = VolumeFilter{FilterVolume: func(vol *v1.Volume) (string, bool) {if vol.Cinder != nil {return vol.Cinder.VolumeID, true}return "", false},FilterPersistentVolume: func(pv *v1.PersistentVolume) (string, bool) {if pv.Spec.Cinder != nil {return pv.Spec.Cinder.VolumeID, true}return "", false},MatchProvisioner: func(sc *storage.StorageClass) (relevant bool) {if sc.Provisioner == csilibplugins.CinderInTreePluginName {return true}return false},IsMigrated: func(csiNode *storage.CSINode) bool {return isCSIMigrationOn(csiNode, csilibplugins.CinderInTreePluginName)}, }

可以看到，filer其實(shí)是通過對元數(shù)據(jù)的字段(vol.Cinder，pv.Spec.Cinder，sc.Provisioner)等進(jìn)行判斷來分辨是不是本類型的volume的。

如果不是intree的，可以通過其它的字段來進(jìn)行過濾

Flexvolume的我們可以用這些字段：

vol.FlexVolume.Driver，pv.Spec.FlexVolume.Driver

CSI的我們可以用這些字段：

vol.CSI.Driver,pv.Spec.CSI.Driver

2.1.3 核心方法Filter

// Filter invoked at the filter extension point. func (pl *nonCSILimits) Filter(ctx context.Context, _ *framework.CycleState, pod *v1.Pod, nodeInfo *framework.NodeInfo) *framework.Status {// If a pod doesn't have any volume attached to it, the predicate will always be true.// Thus we make a fast path for it, to avoid unnecessary computations in this case.if len(pod.Spec.Volumes) == 0 {return nil}newVolumes := make(map[string]bool)if err := pl.filterVolumes(pod.Spec.Volumes, pod.Namespace, newVolumes); err != nil {return framework.NewStatus(framework.Error, err.Error())}// quick returnif len(newVolumes) == 0 {return nil}node := nodeInfo.Node()if node == nil {return framework.NewStatus(framework.Error, fmt.Sprintf("node not found"))}var csiNode *storage.CSINodevar err errorif pl.csiNodeLister != nil {csiNode, err = pl.csiNodeLister.Get(node.Name)if err != nil {// we don't fail here because the CSINode object is only necessary// for determining whether the migration is enabled or notklog.V(5).Infof("Could not get a CSINode object for the node: %v", err)}}// if a plugin has been migrated to a CSI driver, defer to the CSI predicateif pl.filter.IsMigrated(csiNode) {return nil}// count unique volumesexistingVolumes := make(map[string]bool)for _, existingPod := range nodeInfo.Pods {if err := pl.filterVolumes(existingPod.Pod.Spec.Volumes, existingPod.Pod.Namespace, existingVolumes); err != nil {return framework.NewStatus(framework.Error, err.Error())}}numExistingVolumes := len(existingVolumes)// filter out already-mounted volumesfor k := range existingVolumes {delete(newVolumes, k)}numNewVolumes := len(newVolumes)maxAttachLimit := pl.maxVolumeFunc(node)volumeLimits := volumeLimits(nodeInfo)if maxAttachLimitFromAllocatable, ok := volumeLimits[pl.volumeLimitKey]; ok {maxAttachLimit = int(maxAttachLimitFromAllocatable)}if numExistingVolumes+numNewVolumes > maxAttachLimit {return framework.NewStatus(framework.Unschedulable, ErrReasonMaxVolumeCountExceeded)}if nodeInfo != nil && nodeInfo.TransientInfo != nil && utilfeature.DefaultFeatureGate.Enabled(features.BalanceAttachedNodeVolumes) {nodeInfo.TransientInfo.TransientLock.Lock()defer nodeInfo.TransientInfo.TransientLock.Unlock()nodeInfo.TransientInfo.TransNodeInfo.AllocatableVolumesCount = maxAttachLimit - numExistingVolumesnodeInfo.TransientInfo.TransNodeInfo.RequestedVolumes = numNewVolumes}return nil }

接下來看一下核心代碼的邏輯

最開始是一些校驗(yàn)，首先校驗(yàn)這個pod有沒有使用volume，再校驗(yàn)這個pod使用的volume是什么類型的volume，然后由對應(yīng)類型的volume的filter進(jìn)行處理。

看一下最核心的處理流程

numExistingVolumes := len(existingVolumes)// filter out already-mounted volumesfor k := range existingVolumes {delete(newVolumes, k)}numNewVolumes := len(newVolumes)maxAttachLimit := pl.maxVolumeFunc(node)volumeLimits := volumeLimits(nodeInfo)if maxAttachLimitFromAllocatable, ok := volumeLimits[pl.volumeLimitKey]; ok {maxAttachLimit = int(maxAttachLimitFromAllocatable)}if numExistingVolumes+numNewVolumes > maxAttachLimit {return framework.NewStatus(framework.Unschedulable, ErrReasonMaxVolumeCountExceeded)}if nodeInfo != nil && nodeInfo.TransientInfo != nil && utilfeature.DefaultFeatureGate.Enabled(features.BalanceAttachedNodeVolumes) {nodeInfo.TransientInfo.TransientLock.Lock()defer nodeInfo.TransientInfo.TransientLock.Unlock()nodeInfo.TransientInfo.TransNodeInfo.AllocatableVolumesCount = maxAttachLimit - numExistingVolumesnodeInfo.TransientInfo.TransNodeInfo.RequestedVolumes = numNewVolumes}return?nil

邏輯非常清楚

numExistingVolumes 表示該節(jié)點(diǎn)上已經(jīng)有的volume數(shù)

numNewVolumes是該pod需要創(chuàng)建的volume數(shù)

maxAttachLimit 是該類型的volume在節(jié)點(diǎn)上所能創(chuàng)建的最大volume數(shù)

如果 numExistingVolumes+numNewVolumes > maxAttachLimit 則該節(jié)點(diǎn)不可調(diào)度

如果可以調(diào)度，則return nil，然后釋放資源

到此，預(yù)選階段的工作做完。

這里還需要注意一個地方

nodeInfo.TransientInfo.TransNodeInfo.AllocatableVolumesCount = maxAttachLimit - numExistingVolumesnodeInfo.TransientInfo.TransNodeInfo.RequestedVolumes?=?numNewVolumes

AllocatableVolumesCount，RequestedVolumes這兩個變量的賦值是給后面進(jìn)行優(yōu)選打分的時候用的

2.2 磁盤優(yōu)選（Priorities）

優(yōu)選打分的條件也很多，這里我們只處理跟存儲相關(guān)的，具體源碼在

pkg\scheduler\framework\plugins\noderesources\resource_allocation.go

pkg\scheduler\framework\plugins\noderesources\least_allocated.go

2.2.1 resource_allocation

resource_allocation 字體意思就可以理解，就是資源分配，而資源分配有多種方，比如least_allocated，most_allocated，requested_to_capacity_ratio都是資源分配的方法，而resource_allocation只是提供一個score方法，代碼如下:

func (r *resourceAllocationScorer) score(pod *v1.Pod,nodeInfo *framework.NodeInfo) (int64, *framework.Status) {node := nodeInfo.Node()if node == nil {return 0, framework.NewStatus(framework.Error, "node not found")}if r.resourceToWeightMap == nil {return 0, framework.NewStatus(framework.Error, "resources not found")}requested := make(resourceToValueMap, len(r.resourceToWeightMap))allocatable := make(resourceToValueMap, len(r.resourceToWeightMap))for resource := range r.resourceToWeightMap {allocatable[resource], requested[resource] = calculateResourceAllocatableRequest(nodeInfo, pod, resource)}var score int64// Check if the pod has volumes and this could be added to scorer function for balanced resource allocation.if len(pod.Spec.Volumes) > 0 && utilfeature.DefaultFeatureGate.Enabled(features.BalanceAttachedNodeVolumes) && nodeInfo.TransientInfo != nil {score = r.scorer(requested, allocatable, true, nodeInfo.TransientInfo.TransNodeInfo.RequestedVolumes, nodeInfo.TransientInfo.TransNodeInfo.AllocatableVolumesCount)} else {score = r.scorer(requested, allocatable, false, 0, 0)}if klog.V(10).Enabled() {if len(pod.Spec.Volumes) > 0 && utilfeature.DefaultFeatureGate.Enabled(features.BalanceAttachedNodeVolumes) && nodeInfo.TransientInfo != nil {klog.Infof("%v -> %v: %v, map of allocatable resources %v, map of requested resources %v , allocatable volumes %d, requested volumes %d, score %d",pod.Name, node.Name, r.Name,allocatable, requested, nodeInfo.TransientInfo.TransNodeInfo.AllocatableVolumesCount,nodeInfo.TransientInfo.TransNodeInfo.RequestedVolumes,score,)} else {klog.Infof("%v -> %v: %v, map of allocatable resources %v, map of requested resources %v ,score %d,",pod.Name, node.Name, r.Name,allocatable, requested, score,)}}return score, nil }

核心代碼就一句

if len(pod.Spec.Volumes) > 0 && utilfeature.DefaultFeatureGate.Enabled(features.BalanceAttachedNodeVolumes) && nodeInfo.TransientInfo != nil {score = r.scorer(requested, allocatable, true, nodeInfo.TransientInfo.TransNodeInfo.RequestedVolumes, nodeInfo.TransientInfo.TransNodeInfo.AllocatableVolumesCount)} else {score = r.scorer(requested, allocatable, false, 0, 0)}

如果pod有volume并且BalanceAttachedNodeVolumes這個feature打開了，并且節(jié)點(diǎn)有TransientInfo

那么就走存儲相關(guān)的打分，否則就不走。

我們看一下r.scorer的參數(shù)的后兩個，就是我們在預(yù)選階段最后賦值的兩個參數(shù)

AllocatableVolumesCount 表示還可以創(chuàng)建的volume數(shù)量

RequestedVolumes 表示該pod需要的volume數(shù)量

2.2.2 least_allocated

least_allocated代碼最少資源優(yōu)先，也就是節(jié)點(diǎn)上資源越少，分越高

直接看源碼

type resourceAllocationScorer struct {Name stringscorer func(requested, allocable resourceToValueMap, includeVolumes bool, requestedVolumes int, allocatableVolumes int) int64resourceToWeightMap resourceToWeightMap }func leastResourceScorer(resToWeightMap resourceToWeightMap) func(resourceToValueMap, resourceToValueMap, bool, int, int) int64 {return func(requested, allocable resourceToValueMap, includeVolumes bool, requestedVolumes int, allocatableVolumes int) int64 {var nodeScore, weightSum int64for resource, weight := range resToWeightMap {resourceScore := leastRequestedScore(requested[resource], allocable[resource])nodeScore += resourceScore * weightweightSum += weight}return nodeScore / weightSum} }// The unused capacity is calculated on a scale of 0-MaxNodeScore // 0 being the lowest priority and `MaxNodeScore` being the highest. // The more unused resources the higher the score is. func leastRequestedScore(requested, capacity int64) int64 {if capacity == 0 {return 0}if requested > capacity {return 0}return ((capacity - requested) * int64(framework.MaxNodeScore)) / capacity }

我們來看核心算法

capacity表示還可以剩余資源數(shù)量

requested表示該pod需求資源數(shù)量

比如：capacity=10，requested=3，framework.MaxNodeScorel默認(rèn)是100

那么得分就是 (10-3)*100/10=70

但是我們看到 leastResourceScorer中并沒有引用存儲的部分，所以我們可以手動添加上

if includeVolumes && allocatableVolumes - requestedVolumes > 0 && allocatableVolumes > 0 {nodeScore += int64(((allocatableVolumes - requestedVolumes) * int(framework.MaxNodeScore)) / allocatableVolumes)weightSum += 1}

至此，kube-scheduler存儲部分代碼解讀。

往期推薦

從 40% 跌至 4%，“糊”了的 Firefox 還能重回巔峰嗎？

Gartner 發(fā)布 2022 年汽車行業(yè)五大技術(shù)趨勢

使用這個庫，讓你的服務(wù)操作 Redis 速度飛起

漫畫：什么是“低代碼”開發(fā)平臺？

點(diǎn)分享

點(diǎn)收藏

點(diǎn)點(diǎn)贊

點(diǎn)在看

總結(jié)

以上是生活随笔為你收集整理的kube-scheduler 磁盘调度源码分析的全部內(nèi)容，希望文章能夠幫你解決所遇到的問題。

如果覺得生活随笔網(wǎng)站內(nèi)容還不錯，歡迎將生活随笔推薦給好友。