Chapter 17. Scheduler

The available nodes are filtered based on the constraints or requirements specified. This is done by running each of the nodes through the list of filter functions called 'predicates'.

This is achieved by passing each node through a series of 'priority' functions that assign it a score between 0 - 10, with 0 indicating a bad fit and 10 indicating a good fit to host the pod. The scheduler configuration can also take in a simple "weight" (positive numeric value) for each priority function. The node score provided by each priority function is multiplied by the "weight" (default weight is 1) and then combined by just adding the scores for each node provided by all the priority functions. This weight attribute can be used by administrators to give higher importance to some priority functions.

The nodes are sorted based on their scores and the node with the highest score is selected to host the pod. If multiple nodes have the same high score, then one of them is selected at random.

These predicates do not take any configuration parameters or inputs from the user. These are specified in the scheduler configuration using their exact name.

PodFitsPorts deems a node to be fit for hosting a pod based on the absence of port conflicts.

{"name" : "PodFitsPorts"}

PodFitsResources determines a fit based on resource availability. The nodes can declare their resource capacities and then pods can specify what resources they require. Fit is based on requested, rather than used resources.

{"name" : "PodFitsResources"}

NoDiskConflict determines fit based on non-conflicting disk volumes. It evaluates if a pod can fit due to the volumes it requests, and those that are already mounted. It is GCE PD, Amazon EBS, and Ceph RBD specific. Only Persistent Volume Claims for those supported types are checked. Persistent Volumes added directly to pods are not evaluated and are not constrained by this policy.

{"name" : "NoDiskConflict"}

MatchNodeSelector determines fit based on node selector query that is defined in the pod.

{"name" : "MatchNodeSelector"}

HostName determines fit based on the presence of the Host parameter and a string match with the name of the host.

{"name" : "HostName"}

These predicates can be configured by the user to tweak their functioning. They can be given any user-defined name. The type of the predicate is identified by the argument that they take. Since these are configurable, multiple predicates of the same type (but different configuration parameters) can be combined as long as their user-defined names are different.

ServiceAffinity filters out nodes that do not belong to the specified topological level defined by the provided labels. This predicate takes in a list of labels and ensures affinity within the nodes (that have the same label values) for pods belonging to the same service. If the pod specifies a value for the labels in its NodeSelector, then the nodes matching those labels are the ones where the pod is scheduled. If the pod does not specify the labels in its NodeSelector, then the first pod can be placed on any node based on availability and all subsequent pods of the service will be scheduled on nodes that have the same label values.

{"name" : "Zone", "argument" : {"serviceAffinity" : {"labels" : ["zone"]}}}

LabelsPresence checks whether a particular node has a certain label defined or not, regardless of its value. Matching by label can be useful, for example, where nodes have their physical location or status defined by labels.

{"name" : "RequireRegion", "argument" : {"labelsPresence" : {"labels" : ["region"], "presence" : true}}}

These priority functions do not take any configuration parameters or inputs from the user. These are specified in the scheduler configuration using their exact name as well as the weight.

LeastRequestedPriority favors nodes with fewer requested resources. It calculates the percentage of memory and CPU requested by pods scheduled on the node, and prioritizes nodes that have the highest available/remaining capacity.

{"name" : "LeastRequestedPriority", "weight" : 1}

BalancedResourceAllocation favors nodes with balanced resource usage rate. It calculates the difference between the consumed CPU and memory as a fraction of capacity, and prioritizes the nodes based on how close the two metrics are to each other. This should always be used together with LeastRequestedPriority.

{"name" : "BalancedResourceAllocation", "weight" : 1}

ServiceSpreadingPriority spreads pods by minimizing the number of pods belonging to the same service onto the same machine.

{"name" : "ServiceSpreadingPriority", "weight" : 1}

EqualPriority gives an equal weight of one to all nodes, if no priority configs are provided. It is not required/recommended outside of testing.

{"name" : "EqualPriority", "weight" : 1}

These priority functions can be configured by the user by providing certain parameters. They can be given any user-defined name. The type of the priority function is identified by the argument that they take. Since these are configurable, multiple priority functions of the same type (but different configuration parameters) can be combined as long as their user-defined names are different.

ServiceAntiAffinity takes a label and ensures a good spread of the pods belonging to the same service across the group of nodes based on the label values. It gives the same score to all nodes that have the same value for the specified label. It gives a higher score to nodes within a group with the least concentration of pods.

{"name" : "RackSpread", "weight" : 1, "argument" : {"serviceAntiAffinity" : {"label" : "rack"}}}

LabelPreference prefers nodes that have a particular label defined or not, regardless of its value.

{"name" : "RackPreferred", "weight" : 1, "argument" : {"labelPreference" : {"label" : "rack"}}}

The default scheduler policy includes the following predicates:

The default scheduler policy includes the following priority functions. Each of the priority function has a weight of '1' applied to it:

The scheduler policy is defined in a file on the master, named /etc/origin/master/scheduler.json by default, unless overridden by the kubernetesMasterConfig.schedulerConfigFile field in the master configuration file.

To modify the scheduler policy:

Administrators can define multiple topological levels for their infrastructure (nodes). This is done by specifying labels on nodes (e.g., region=r1, zone=z1, rack=s1). These label names have no particular meaning and administrators are free to name their infrastructure levels anything (eg, city/building/room). Also, administrators can define any number of levels for their infrastructure topology, with three levels usually being adequate (eg. regions zones racks). Lastly, administrators can specify affinity and anti-affinity rules at each of these levels in any combination.

Administrators should be able to configure the scheduler to specify affinity at any topological level, or even at multiple levels. Affinity at a particular level indicates that all pods that belong to the same service will be scheduled onto nodes that belong to the same level. This handles any latency requirements of applications by allowing administrators to ensure that peer pods do not end up being too geographically separated. If no node is available within the same affinity group to host the pod, then the pod will not get scheduled.

Administrators should be able to configure the scheduler to specify anti-affinity at any topological level, or even at multiple levels. Anti-Affinity (or 'spread') at a particular level indicates that all pods that belong to the same service will be spread across nodes that belong to that level. This ensures that the application is well spread for high availability purposes. The scheduler will try to balance the service pods across all applicable nodes as evenly as possible.

The scheduler functionality can be enhanced by adding new predicates and priority functions. They can either be contributed upstream or maintained separately. These predicates and priority functions would need to be registered with the scheduler factory and then specified in the scheduler policy file.

Since the scheduler is a plug-in, it can be replaced in favor of an alternate implementation. The scheduler code has a clean separation that watches new pods as they get created and identifies the most suitable node to host them. It then creates bindings (pod to node bindings) for the pods using the master API.

Ensure a pod is deployed onto only a specified node host by assigning it a label and specifying this in the nodeName setting in a pod configuration.

Using nodeSelector in a pod configuration, you can ensure that pods are only placed onto nodes with specific labels.