github/kubernetes
1
0
Fork 0
mirror of https://github.com/k3s-io/kubernetes.git synced 2025-11-23 12:29:01 +00:00
Code Issues Projects Releases Wiki Activity

Merge pull request #50949 from bsalamat/preemption_eviction

Browse Source 
Automatic merge from submit-queue

Add pod preemption to the scheduler

**What this PR does / why we need it**:
This is the last of a series of PRs to add priority-based preemption to the scheduler. This PR connects the preemption logic to the scheduler workflow.

**Which issue this PR fixes** *(optional, in `fixes #<issue number>(, fixes #<issue_number>, ...)` format, will close that issue when PR gets merged)*: fixes #48646

**Special notes for your reviewer**:
This PR includes other PRs which are under review (#50805, #50405, #50190). All the new code is located in 43627afdf9.

**Release note**:

```release-note
Add priority-based preemption to the scheduler.
```

ref/ #47604

/assign @davidopp 

@kubernetes/sig-scheduling-pr-reviews
This commit is contained in:
Kubernetes Submit Queue
2017-09-08 14:19:42 -07:00
committed by GitHub
parent ed154988c5 c0b718373b
commit f695a3120a
34 changed files with 1900 additions and 91 deletions
Download Patch File Download Diff File Expand all files Collapse all files

355
plugin/pkg/scheduler/core/generic_scheduler.go
View File

@@ -18,6 +18,7 @@ package core
import (
"fmt"
"math"
"sort"
"strings"
"sync"
@@ -32,6 +33,7 @@ import (
"k8s.io/kubernetes/plugin/pkg/scheduler/algorithm/predicates"
schedulerapi "k8s.io/kubernetes/plugin/pkg/scheduler/api"
"k8s.io/kubernetes/plugin/pkg/scheduler/schedulercache"
"k8s.io/kubernetes/plugin/pkg/scheduler/util"
"github.com/golang/glog"
)
@@ -45,7 +47,14 @@ type FitError struct {
var ErrNoNodesAvailable = fmt.Errorf("no nodes available to schedule pods")
const NoNodeAvailableMsg = "No nodes are available that match all of the following predicates"
const (
NoNodeAvailableMsg = "No nodes are available that match all of the predicates"
// NominatedNodeAnnotationKey is used to annotate a pod that has preempted other pods.
// The scheduler uses the annotation to find that the pod shouldn't preempt more pods
// when it gets to the head of scheduling queue again.
// See podEligibleToPreemptOthers() for more information.
NominatedNodeAnnotationKey = "NominatedNodeName"
)
// Error returns detailed information of why the pod failed to fit on each node
func (f *FitError) Error() string {
@@ -73,7 +82,7 @@ type genericScheduler struct {
equivalenceCache *EquivalenceCache
predicates map[string]algorithm.FitPredicate
priorityMetaProducer algorithm.MetadataProducer
predicateMetaProducer algorithm.MetadataProducer
predicateMetaProducer algorithm.PredicateMetadataProducer
prioritizers []algorithm.PriorityConfig
extenders []algorithm.SchedulerExtender
pods algorithm.PodLister
@@ -159,6 +168,65 @@ func (g *genericScheduler) selectHost(priorityList schedulerapi.HostPriorityList
return priorityList[ix].Host, nil
}
// preempt finds nodes with pods that can be preempted to make room for "pod" to
// schedule. It chooses one of the nodes and preempts the pods on the node and
// returns the node and the list of preempted pods if such a node is found.
// TODO(bsalamat): Add priority-based scheduling. More info: today one or more
// pending pods (different from the pod that triggered the preemption(s)) may
// schedule into some portion of the resources freed up by the preemption(s)
// before the pod that triggered the preemption(s) has a chance to schedule
// there, thereby preventing the pod that triggered the preemption(s) from
// scheduling. Solution is given at:
// https://github.com/kubernetes/community/blob/master/contributors/design-proposals/pod-preemption.md#preemption-mechanics
func (g *genericScheduler) Preempt(pod *v1.Pod, nodeLister algorithm.NodeLister, scheduleErr error) (*v1.Node, []*v1.Pod, error) {
// Scheduler may return various types of errors. Consider preemption only if
// the error is of type FitError.
fitError, ok := scheduleErr.(*FitError)
if !ok || fitError == nil {
return nil, nil, nil
}
err := g.cache.UpdateNodeNameToInfoMap(g.cachedNodeInfoMap)
if err != nil {
return nil, nil, err
}
if !podEligibleToPreemptOthers(pod, g.cachedNodeInfoMap) {
glog.V(5).Infof("Pod %v is not eligible for more preemption.", pod.Name)
return nil, nil, nil
}
allNodes, err := nodeLister.List()
if err != nil {
return nil, nil, err
}
if len(allNodes) == 0 {
return nil, nil, ErrNoNodesAvailable
}
potentialNodes := nodesWherePreemptionMightHelp(pod, allNodes, fitError.FailedPredicates)
if len(potentialNodes) == 0 {
glog.V(3).Infof("Preemption will not help schedule pod %v on any node.", pod.Name)
return nil, nil, nil
}
nodeToPods, err := selectNodesForPreemption(pod, g.cachedNodeInfoMap, potentialNodes, g.predicates, g.predicateMetaProducer)
if err != nil {
return nil, nil, err
}
for len(nodeToPods) > 0 {
node := pickOneNodeForPreemption(nodeToPods)
if node == nil {
return nil, nil, err
}
passes, pErr := nodePassesExtendersForPreemption(pod, node.Name, nodeToPods[node], g.cachedNodeInfoMap, g.extenders)
if passes && pErr == nil {
return node, nodeToPods[node], err
}
if pErr != nil {
glog.Errorf("Error occurred while checking extenders for preemption on node %v: %v", node, pErr)
}
// Remove the node from the map and try to pick a different node.
delete(nodeToPods, node)
}
return nil, nil, err
}
// Filters the nodes to find the ones that fit based on the given predicate functions
// Each node is passed through the predicate functions to determine if it is a fit
func findNodesThatFit(
@@ -167,7 +235,7 @@ func findNodesThatFit(
nodes []*v1.Node,
predicateFuncs map[string]algorithm.FitPredicate,
extenders []algorithm.SchedulerExtender,
metadataProducer algorithm.MetadataProducer,
metadataProducer algorithm.PredicateMetadataProducer,
ecache *EquivalenceCache,
) ([]*v1.Node, FailedPredicateMap, error) {
var filtered []*v1.Node
@@ -232,7 +300,7 @@ func findNodesThatFit(
}
// Checks whether node with a given name and NodeInfo satisfies all predicateFuncs.
func podFitsOnNode(pod *v1.Pod, meta interface{}, info *schedulercache.NodeInfo, predicateFuncs map[string]algorithm.FitPredicate,
func podFitsOnNode(pod *v1.Pod, meta algorithm.PredicateMetadata, info *schedulercache.NodeInfo, predicateFuncs map[string]algorithm.FitPredicate,
ecache *EquivalenceCache) (bool, []algorithm.PredicateFailureReason, error) {
var (
equivalenceHash uint64
@@ -422,11 +490,288 @@ func EqualPriorityMap(_ *v1.Pod, _ interface{}, nodeInfo *schedulercache.NodeInf
}, nil
}
// pickOneNodeForPreemption chooses one node among the given nodes. It assumes
// pods in each map entry are ordered by decreasing priority.
// It picks a node based on the following criteria:
// 1. A node with minimum highest priority victim is picked.
// 2. Ties are broken by sum of priorities of all victims.
// 3. If there are still ties, node with the minimum number of victims is picked.
// 4. If there are still ties, the first such node is picked (sort of randomly).
//TODO(bsalamat): Try to reuse the "nodeScore" slices in order to save GC time.
func pickOneNodeForPreemption(nodesToPods map[*v1.Node][]*v1.Pod) *v1.Node {
type nodeScore struct {
node *v1.Node
highestPriority int32
sumPriorities int64
numPods int
}
if len(nodesToPods) == 0 {
return nil
}
minHighestPriority := int32(math.MaxInt32)
minPriorityScores := []*nodeScore{}
for node, pods := range nodesToPods {
if len(pods) == 0 {
// We found a node that doesn't need any preemption. Return it!
// This should happen rarely when one or more pods are terminated between
// the time that scheduler tries to schedule the pod and the time that
// preemption logic tries to find nodes for preemption.
return node
}
// highestPodPriority is the highest priority among the victims on this node.
highestPodPriority := util.GetPodPriority(pods[0])
if highestPodPriority < minHighestPriority {
minHighestPriority = highestPodPriority
minPriorityScores = nil
}
if highestPodPriority == minHighestPriority {
minPriorityScores = append(minPriorityScores, &nodeScore{node: node, highestPriority: highestPodPriority, numPods: len(pods)})
}
}
if len(minPriorityScores) == 1 {
return minPriorityScores[0].node
}
// There are a few nodes with minimum highest priority victim. Find the
// smallest sum of priorities.
minSumPriorities := int64(math.MaxInt64)
minSumPriorityScores := []*nodeScore{}
for _, nodeScore := range minPriorityScores {
var sumPriorities int64
for _, pod := range nodesToPods[nodeScore.node] {
// We add MaxInt32+1 to all priorities to make all of them >= 0. This is
// needed so that a node with a few pods with negative priority is not
// picked over a node with a smaller number of pods with the same negative
// priority (and similar scenarios).
sumPriorities += int64(util.GetPodPriority(pod)) + int64(math.MaxInt32+1)
}
if sumPriorities < minSumPriorities {
minSumPriorities = sumPriorities
minSumPriorityScores = nil
}
nodeScore.sumPriorities = sumPriorities
if sumPriorities == minSumPriorities {
minSumPriorityScores = append(minSumPriorityScores, nodeScore)
}
}
if len(minSumPriorityScores) == 1 {
return minSumPriorityScores[0].node
}
// There are a few nodes with minimum highest priority victim and sum of priorities.
// Find one with the minimum number of pods.
minNumPods := math.MaxInt32
minNumPodScores := []*nodeScore{}
for _, nodeScore := range minSumPriorityScores {
if nodeScore.numPods < minNumPods {
minNumPods = nodeScore.numPods
minNumPodScores = nil
}
if nodeScore.numPods == minNumPods {
minNumPodScores = append(minNumPodScores, nodeScore)
}
}
// At this point, even if there are more than one node with the same score,
// return the first one.
if len(minNumPodScores) > 0 {
return minNumPodScores[0].node
}
glog.Errorf("Error in logic of node scoring for preemption. We should never reach here!")
return nil
}
// selectNodesForPreemption finds all the nodes with possible victims for
// preemption in parallel.
func selectNodesForPreemption(pod *v1.Pod,
nodeNameToInfo map[string]*schedulercache.NodeInfo,
potentialNodes []*v1.Node,
predicates map[string]algorithm.FitPredicate,
metadataProducer algorithm.PredicateMetadataProducer,
) (map[*v1.Node][]*v1.Pod, error) {
nodeNameToPods := map[*v1.Node][]*v1.Pod{}
var resultLock sync.Mutex
// We can use the same metadata producer for all nodes.
meta := metadataProducer(pod, nodeNameToInfo)
checkNode := func(i int) {
nodeName := potentialNodes[i].Name
var metaCopy algorithm.PredicateMetadata
if meta != nil {
metaCopy = meta.ShallowCopy()
}
pods, fits := selectVictimsOnNode(pod, metaCopy, nodeNameToInfo[nodeName], predicates)
if fits {
resultLock.Lock()
nodeNameToPods[potentialNodes[i]] = pods
resultLock.Unlock()
}
}
workqueue.Parallelize(16, len(potentialNodes), checkNode)
return nodeNameToPods, nil
}
func nodePassesExtendersForPreemption(
pod *v1.Pod,
nodeName string,
victims []*v1.Pod,
nodeNameToInfo map[string]*schedulercache.NodeInfo,
extenders []algorithm.SchedulerExtender) (bool, error) {
// If there are any extenders, run them and filter the list of candidate nodes.
if len(extenders) == 0 {
return true, nil
}
// Remove the victims from the corresponding nodeInfo and send nodes to the
// extenders for filtering.
originalNodeInfo := nodeNameToInfo[nodeName]
nodeInfoCopy := nodeNameToInfo[nodeName].Clone()
for _, victim := range victims {
nodeInfoCopy.RemovePod(victim)
}
nodeNameToInfo[nodeName] = nodeInfoCopy
defer func() { nodeNameToInfo[nodeName] = originalNodeInfo }()
filteredNodes := []*v1.Node{nodeInfoCopy.Node()}
for _, extender := range extenders {
var err error
var failedNodesMap map[string]string
filteredNodes, failedNodesMap, err = extender.Filter(pod, filteredNodes, nodeNameToInfo)
if err != nil {
return false, err
}
if _, found := failedNodesMap[nodeName]; found || len(filteredNodes) == 0 {
return false, nil
}
}
return true, nil
}
// selectVictimsOnNode finds minimum set of pods on the given node that should
// be preempted in order to make enough room for "pod" to be scheduled. The
// minimum set selected is subject to the constraint that a higher-priority pod
// is never preempted when a lower-priority pod could be (higher/lower relative
// to one another, not relative to the preemptor "pod").
// The algorithm first checks if the pod can be scheduled on the node when all the
// lower priority pods are gone. If so, it sorts all the lower priority pods by
// their priority and starts from the highest priority one, tries to keep as
// many of them as possible while checking that the "pod" can still fit on the node.
// NOTE: This function assumes that it is never called if "pod" cannot be scheduled
// due to pod affinity, node affinity, or node anti-affinity reasons. None of
// these predicates can be satisfied by removing more pods from the node.
// TODO(bsalamat): Add support for PodDisruptionBudget.
func selectVictimsOnNode(
pod *v1.Pod,
meta algorithm.PredicateMetadata,
nodeInfo *schedulercache.NodeInfo,
fitPredicates map[string]algorithm.FitPredicate) ([]*v1.Pod, bool) {
potentialVictims := util.SortableList{CompFunc: util.HigherPriorityPod}
nodeInfoCopy := nodeInfo.Clone()
removePod := func(rp *v1.Pod) {
nodeInfoCopy.RemovePod(rp)
if meta != nil {
meta.RemovePod(rp)
}
}
addPod := func(ap *v1.Pod) {
nodeInfoCopy.AddPod(ap)
if meta != nil {
meta.AddPod(ap, nodeInfoCopy)
}
}
// As the first step, remove all the lower priority pods from the node and
// check if the given pod can be scheduled.
podPriority := util.GetPodPriority(pod)
for _, p := range nodeInfoCopy.Pods() {
if util.GetPodPriority(p) < podPriority {
potentialVictims.Items = append(potentialVictims.Items, p)
removePod(p)
}
}
potentialVictims.Sort()
// If the new pod does not fit after removing all the lower priority pods,
// we are almost done and this node is not suitable for preemption. The only condition
// that we should check is if the "pod" is failing to schedule due to pod affinity
// failure.
// TODO(bsalamat): Consider checking affinity to lower priority pods if feasible with reasonable performance.
if fits, _, err := podFitsOnNode(pod, meta, nodeInfoCopy, fitPredicates, nil); !fits {
if err != nil {
glog.Warningf("Encountered error while selecting victims on node %v: %v", nodeInfo.Node().Name, err)
}
return nil, false
}
victims := []*v1.Pod{}
// Try to reprieve as many pods as possible starting from the highest priority one.
for _, p := range potentialVictims.Items {
lpp := p.(*v1.Pod)
addPod(lpp)
if fits, _, _ := podFitsOnNode(pod, meta, nodeInfoCopy, fitPredicates, nil); !fits {
removePod(lpp)
victims = append(victims, lpp)
glog.V(5).Infof("Pod %v is a potential preemption victim on node %v.", lpp.Name, nodeInfo.Node().Name)
}
}
return victims, true
}
// nodesWherePreemptionMightHelp returns a list of nodes with failed predicates
// that may be satisfied by removing pods from the node.
func nodesWherePreemptionMightHelp(pod *v1.Pod, nodes []*v1.Node, failedPredicatesMap FailedPredicateMap) []*v1.Node {
potentialNodes := []*v1.Node{}
for _, node := range nodes {
unresolvableReasonExist := false
failedPredicates, found := failedPredicatesMap[node.Name]
// If we assume that scheduler looks at all nodes and populates the failedPredicateMap
// (which is the case today), the !found case should never happen, but we'd prefer
// to rely less on such assumptions in the code when checking does not impose
// significant overhead.
for _, failedPredicate := range failedPredicates {
switch failedPredicate {
case
predicates.ErrNodeSelectorNotMatch,
predicates.ErrPodNotMatchHostName,
predicates.ErrTaintsTolerationsNotMatch,
predicates.ErrNodeLabelPresenceViolated,
predicates.ErrNodeNotReady,
predicates.ErrNodeNetworkUnavailable,
predicates.ErrNodeUnschedulable,
predicates.ErrNodeUnknownCondition:
unresolvableReasonExist = true
break
// TODO(bsalamat): Please add affinity failure cases once we have specific affinity failure errors.
}
}
if !found || !unresolvableReasonExist {
glog.V(3).Infof("Node %v is a potential node for preemption.", node.Name)
potentialNodes = append(potentialNodes, node)
}
}
return potentialNodes
}
// podEligibleToPreemptOthers determines whether this pod should be considered
// for preempting other pods or not. If this pod has already preempted other
// pods and those are in their graceful termination period, it shouldn't be
// considered for preemption.
// We look at the node that is nominated for this pod and as long as there are
// terminating pods on the node, we don't consider this for preempting more pods.
// TODO(bsalamat): Revisit this algorithm once scheduling by priority is added.
func podEligibleToPreemptOthers(pod *v1.Pod, nodeNameToInfo map[string]*schedulercache.NodeInfo) bool {
if nodeName, found := pod.Annotations[NominatedNodeAnnotationKey]; found {
if nodeInfo, found := nodeNameToInfo[nodeName]; found {
for _, p := range nodeInfo.Pods() {
if p.DeletionTimestamp != nil && util.GetPodPriority(p) < util.GetPodPriority(pod) {
// There is a terminating pod on the nominated node.
return false
}
}
}
}
return true
}
func NewGenericScheduler(
cache schedulercache.Cache,
eCache *EquivalenceCache,
predicates map[string]algorithm.FitPredicate,
predicateMetaProducer algorithm.MetadataProducer,
predicateMetaProducer algorithm.PredicateMetadataProducer,
prioritizers []algorithm.PriorityConfig,
priorityMetaProducer algorithm.MetadataProducer,
extenders []algorithm.SchedulerExtender) algorithm.ScheduleAlgorithm {