// MapPodsToMachines obtains a list of pods and pivots that list into a map where the keys are host names
// and the values are the list of pods running on that host.
func MapPodsToMachines(lister algorithm.PodLister) (map[string][]*api.Pod, error) {
machineToPods := map[string][]*api.Pod{}
// TODO: perform more targeted query...
pods, err := lister.List(labels.Everything())
if err != nil {
return map[string][]*api.Pod{}, err
}
for _, scheduledPod := range pods {
host := scheduledPod.Spec.NodeName
machineToPods[host] = append(machineToPods[host], scheduledPod)
}
return machineToPods, nil
}
// CalculateAntiAffinityPriority spreads pods by minimizing the number of pods belonging to the same service
// on machines with the same value for a particular label.
// The label to be considered is provided to the struct (ServiceAntiAffinity).
func (s *ServiceAntiAffinity) CalculateAntiAffinityPriority(pod *api.Pod, podLister algorithm.PodLister, nodeLister algorithm.NodeLister) (algorithm.HostPriorityList, error) {
var nsServicePods []*api.Pod
services, err := s.serviceLister.GetPodServices(pod)
if err == nil {
// just use the first service and get the other pods within the service
// TODO: a separate predicate can be created that tries to handle all services for the pod
selector := labels.SelectorFromSet(services[0].Spec.Selector)
pods, err := podLister.List(selector)
if err != nil {
return nil, err
}
// consider only the pods that belong to the same namespace
for _, nsPod := range pods {
if nsPod.Namespace == pod.Namespace {
nsServicePods = append(nsServicePods, nsPod)
}
}
}
nodes, err := nodeLister.List()
if err != nil {
return nil, err
}
// separate out the nodes that have the label from the ones that don't
otherNodes := []string{}
labeledNodes := map[string]string{}
for _, node := range nodes.Items {
if labels.Set(node.Labels).Has(s.label) {
label := labels.Set(node.Labels).Get(s.label)
labeledNodes[node.Name] = label
} else {
otherNodes = append(otherNodes, node.Name)
}
}
podCounts := map[string]int{}
for _, pod := range nsServicePods {
label, exists := labeledNodes[pod.Spec.NodeName]
if !exists {
continue
}
podCounts[label]++
}
numServicePods := len(nsServicePods)
result := []algorithm.HostPriority{}
//score int - scale of 0-10
// 0 being the lowest priority and 10 being the highest
for node := range labeledNodes {
// initializing to the default/max node score of 10
fScore := float32(10)
if numServicePods > 0 {
fScore = 10 * (float32(numServicePods-podCounts[labeledNodes[node]]) / float32(numServicePods))
}
result = append(result, algorithm.HostPriority{Host: node, Score: int(fScore)})
}
// add the open nodes with a score of 0
for _, node := range otherNodes {
result = append(result, algorithm.HostPriority{Host: node, Score: 0})
}
return result, nil
}
// CalculateSpreadPriority spreads pods by minimizing the number of pods belonging to the same service or replication controller. It counts number of pods that run under
// Services or RCs as the pod being scheduled and tries to minimize the number of conflicts. I.e. pushes scheduler towards a Node where there's a smallest number of
// pods which match the same selectors of Services and RCs as current pod.
func (s *SelectorSpread) CalculateSpreadPriority(pod *api.Pod, podLister algorithm.PodLister, nodeLister algorithm.NodeLister) (algorithm.HostPriorityList, error) {
var maxCount int
var nsPods []*api.Pod
selectors := make([]labels.Selector, 0)
services, err := s.serviceLister.GetPodServices(pod)
if err == nil {
for _, service := range services {
selectors = append(selectors, labels.SelectorFromSet(service.Spec.Selector))
}
}
controllers, err := s.controllerLister.GetPodControllers(pod)
if err == nil {
for _, controller := range controllers {
selectors = append(selectors, labels.SelectorFromSet(controller.Spec.Selector))
}
}
if len(selectors) > 0 {
pods, err := podLister.List(labels.Everything())
if err != nil {
return nil, err
}
// consider only the pods that belong to the same namespace
for _, nsPod := range pods {
if nsPod.Namespace == pod.Namespace {
nsPods = append(nsPods, nsPod)
}
}
}
nodes, err := nodeLister.List()
if err != nil {
return nil, err
}
counts := map[string]int{}
if len(nsPods) > 0 {
for _, pod := range nsPods {
matches := false
for _, selector := range selectors {
if selector.Matches(labels.Set(pod.ObjectMeta.Labels)) {
matches = true
break
}
}
if matches {
counts[pod.Spec.NodeName]++
// Compute the maximum number of pods hosted on any node
if counts[pod.Spec.NodeName] > maxCount {
maxCount = counts[pod.Spec.NodeName]
}
}
}
}
result := []algorithm.HostPriority{}
//score int - scale of 0-10
// 0 being the lowest priority and 10 being the highest
for _, node := range nodes.Items {
// initializing to the default/max node score of 10
fScore := float32(10)
if maxCount > 0 {
fScore = 10 * (float32(maxCount-counts[node.Name]) / float32(maxCount))
}
result = append(result, algorithm.HostPriority{Host: node.Name, Score: int(fScore)})
glog.V(10).Infof(
"%v -> %v: SelectorSpreadPriority, Score: (%d)", pod.Name, node.Name, int(fScore),
)
}
return result, nil
}
// CalculateSpreadPriority spreads pods across hosts and zones, considering pods belonging to the same service or replication controller.
// When a pod is scheduled, it looks for services or RCs that match the pod, then finds existing pods that match those selectors.
// It favors nodes that have fewer existing matching pods.
// i.e. it pushes the scheduler towards a node where there's the smallest number of
// pods which match the same service selectors or RC selectors as the pod being scheduled.
// Where zone information is included on the nodes, it favors nodes in zones with fewer existing matching pods.
func (s *SelectorSpread) CalculateSpreadPriority(pod *api.Pod, machinesToPods map[string][]*api.Pod, podLister algorithm.PodLister, nodeLister algorithm.NodeLister) (schedulerapi.HostPriorityList, error) {
var nsPods []*api.Pod
selectors := make([]labels.Selector, 0)
services, err := s.serviceLister.GetPodServices(pod)
if err == nil {
for _, service := range services {
selectors = append(selectors, labels.SelectorFromSet(service.Spec.Selector))
}
}
controllers, err := s.controllerLister.GetPodControllers(pod)
if err == nil {
for _, controller := range controllers {
selectors = append(selectors, labels.SelectorFromSet(controller.Spec.Selector))
}
}
if len(selectors) > 0 {
pods, err := podLister.List(labels.Everything())
if err != nil {
return nil, err
}
// consider only the pods that belong to the same namespace
for _, nsPod := range pods {
if nsPod.Namespace == pod.Namespace {
nsPods = append(nsPods, nsPod)
}
}
}
nodes, err := nodeLister.List()
if err != nil {
return nil, err
}
// Count similar pods by node
countsByNodeName := map[string]int{}
for _, pod := range nsPods {
// When we are replacing a failed pod, we often see the previous deleted version
// while scheduling the replacement. Ignore the previous deleted version for spreading
// purposes (it can still be considered for resource restrictions etc.)
if pod.DeletionTimestamp != nil {
glog.V(2).Infof("skipping pending-deleted pod: %s/%s", pod.Namespace, pod.Name)
continue
}
matches := false
for _, selector := range selectors {
if selector.Matches(labels.Set(pod.ObjectMeta.Labels)) {
matches = true
break
}
}
if !matches {
continue
}
countsByNodeName[pod.Spec.NodeName]++
}
// Aggregate by-node information
// Compute the maximum number of pods hosted on any node
maxCountByNodeName := 0
for _, count := range countsByNodeName {
if count > maxCountByNodeName {
maxCountByNodeName = count
}
}
// Count similar pods by zone, if zone information is present
countsByZone := map[string]int{}
for i := range nodes.Items {
node := &nodes.Items[i]
count, found := countsByNodeName[node.Name]
if !found {
continue
}
zoneId := getZoneKey(node)
if zoneId == "" {
continue
}
countsByZone[zoneId] += count
}
// Aggregate by-zone information
// Compute the maximum number of pods hosted in any zone
haveZones := len(countsByZone) != 0
maxCountByZone := 0
for _, count := range countsByZone {
if count > maxCountByZone {
maxCountByZone = count
}
}
result := []schedulerapi.HostPriority{}
//score int - scale of 0-maxPriority
// 0 being the lowest priority and maxPriority being the highest
for i := range nodes.Items {
node := &nodes.Items[i]
// initializing to the default/max node score of maxPriority
fScore := float32(maxPriority)
if maxCountByNodeName > 0 {
fScore = maxPriority * (float32(maxCountByNodeName-countsByNodeName[node.Name]) / float32(maxCountByNodeName))
}
// If there is zone information present, incorporate it
if haveZones {
zoneId := getZoneKey(node)
if zoneId != "" {
zoneScore := maxPriority * (float32(maxCountByZone-countsByZone[zoneId]) / float32(maxCountByZone))
fScore = (fScore * (1.0 - zoneWeighting)) + (zoneWeighting * zoneScore)
}
}
result = append(result, schedulerapi.HostPriority{Host: node.Name, Score: int(fScore)})
glog.V(10).Infof(
"%v -> %v: SelectorSpreadPriority, Score: (%d)", pod.Name, node.Name, int(fScore),
)
}
return result, nil
}