Skip to main content
Version: v1.5 🚧

GangScheduling

Summary​

This proposal provides Gang mechanism for the scheduler to control pods binding opportunity. User can declare a resource-collection-minimum number, only when assigned-resources reach the given limitation can trigger the binding. We provide Strict and NonStrict to control the resource-accumulation-process by a configuration. We also provide a two-level Gang description for better matching the real scenario, which is different from community.

Motivation​

In AI scenarios, lots of jobs need Gang scheduling. The community have lots of related implements such as Coscheduling or vocalno. We received lots of inspirations in the design process from them.

Compared with competitors​

Coscheduling​

  1. Coscheduling implement a new queue-sort interface and other methods to let one Gang's pods get out of the queue in order as much as possible. If a pod failed to be scheduled, the requests that have been successfully scheduled in this round of Gang scheduling cycle will be rolled back, and the remaining pods waiting for scheduling will be rejected in PreFilter check until this scheduling cycle passed. For example, there is a Gang requires 10 tasks to be scheduled, if first 5 tasks allocated, the 6th task failed to be scheduled, Coscheduling will roll-back first 5 tasks and ignore the remaining 4 tasks in this Gang scheduling cycle. Coscheduling simply use a global time interval to control the Gang scheduling cycle. The first defect is that the uniform time interval will cause some problems. If the time configuration is too long, it will lead to useless waiting; If the time configuration is too short, it will lead to useless scheduling. Secondly, it is very difficult for a large job to meet all resource requests at one time. This mechanism will lead to a very low probability of full resources, and eventually make the job starve to death. We call this process as Strict.

  2. Some jobs have complex Gang requirements. For example, a job has several roles. Each role will have several pods and its own Gang conditions. Jobs also need different roles to form different GangGroups. All pods in a GangGroup can trigger the bind process only after all roles in a GangGroup meet their Gang conditions. The Coscheduling can't meet this requirement.

Goals​

  1. Define API to announce Gang scheduling configuration.

  2. Provides a scheduler plugin to achieve Gang scheduling ability.

Non Goals and Future Work​

  1. Provide ability to solve Gang resource deadlock problems with NonStrict.

Proposal​

Key concept​

Strict and NonStrict​

As mentioned above, in Strict, if a pod failed to be scheduled, the pods that have been successfully scheduled in this scheduling cycle will be rolled back, and the remaining pods waiting for scheduling will be rejected in PreFilter check util this scheduling cycle passed. We call this mode is Strict.

In NonStrict, if a pod failed to be scheduled, it has no impact on any other pod. We will continue to accumulate the allocated pod until the condition of Gang is met. This process is friendly to Gangs with large number of pods, but it will increase the risk of resource deadlock between Gangs. For example, the quota of the quota group is 10(quota will be proposed later), and the user submits three Gangs with 5 pods. Due to various plugin constraints, Gang1\2\3 may allocate resources of 3\3\4 respectively. Since the quota group's quota is full, there will be no new resource scheduling. We call this is resource deadlock of resource Gang. In future proposal, we will try to fix this problem.

GangGroup​

As mentioned above, Some jobs have complex Gang requirements. For example, a job has several roles. Each role will have several pods and its own Gang conditions. Jobs also need different roles to form different GangGroups. All pods in a GangGroup can trigger the bind process only after all roles in a GangGroup meet their Gang conditions. So we introduce GangGroup concept, which allow user to bundle different Gangs together.

After Gang​

It should be noted that, if the resource accumulation conditions of Gang are met, then some pods failed in the process of binding, or some bound pods are preempted\rescheduled, should the constraints of Gang still be effective in the process of resource reallocation? Because the initial purpose of Gang is to require pods to be pulled up at the same time, if some pods have been pulled up, then the subsequent Gang behavior is meaningless. Therefore, when once Gang has been satisfied, all subsequent resource allocations are no longer constrained by Gang rules, and their performance is similar to ordinary pod.

As mentioned above, WaitTime is the max wait time since first pod comes to permit stage. If WaitTime is timeout, scheduler will roll back all assumed pods, update each pod's annotation with gang.scheduling.koordinator.sh/timeout=true, and won't schedule these pods anymore. User should pay attention to this status and delete pods timely.

API​

Definition​

Our original intention is to improve and enhance the ability of the community's original PodGroup, so we will be compatible with the way the community declares the PodGroup. We also provide a lighting way to just use annotations to use Gang feature.

CRD way​

User can use PodGroup CRD in community to declare a gang:

type PodGroup struct {
metav1.TypeMeta `json:",inline"`
metav1.ObjectMeta `json:"metadata,omitempty"`
Spec PodGroupSpec `json:"spec,omitempty"`
Status PodGroupStatus `json:"status,omitempty"`
}
type PodGroupSpec struct {
MinMember int32 `json:"minMember,omitempty"`
MinResources *v1.ResourceList `json:"minResources,omitempty"`

ScheduleTimeoutSeconds *int32 `json:"scheduleTimeoutSeconds,omitempty"`
}

Pod should use pod-group.scheduling.sigs.k8s.io in label to associate with PodGroup.

Also, we introduce some optional definitions as below:

gang.scheduling.koordinator.sh/total-number
gang.scheduling.koordinator.sh/mode
gang.scheduling.koordinator.sh/groups
  • gang.scheduling.koordinator.sh/name indicates the gang's name, it should be emphasized that the name should be in the form of RFC 1123

  • gang.scheduling.koordinator.sh/total-number helps to calculate Gang scheduling cycle in strict mode, you can find more detail in Data-Structure chapter. Default equals to gang.scheduling.koordinator.sh/min-available.

  • gang.scheduling.koordinator.sh/mode determines Strict or NonStrict. Default is Strict.

  • gang.scheduling.koordinator.sh/groups describes GangGroups. Default is empty, which means don't need to form a GangGroup with others, and the gangs in one gangGroup can from different namespaces.

gang.scheduling.koordinator.sh/total-number, gang.scheduling.koordinator.sh/mode, gang.scheduling.koordinator.sh/gang-groups should be found in PodGroup's annotation if needed.

Example​

When user apply a basic gang, the example is as follows:

apiVersion: v1alpha1
kind: PodGroup
metadata:
creationTimestamp: "2022-07-11T18:26:33Z"
name: gang-a
namespace: default
spec:
minMember: 5
minResources:
cpu: "5"
memory: "2048Mi"
scheduleTimeoutSeconds: 600

Let's assume a job has two roles: A and B, each role has several pods. podA belongs to roleA, podB belongs to roleB. roleA and roleB belongs to one GangGroup, the example is as follows:

apiVersion: v1alpha1
kind: PodGroup
metadata:
creationTimestamp: "2022-07-11T18:26:33Z"
name: gang-a
namespace: namespaceA
annotations:
gang.scheduling.koordinator.sh/total-number: 5
gang.scheduling.koordinator.sh/mode: Strict
gang.scheduling.koordinator.sh/groups: ["namespaceA/gang-a", "namespaceB/gang-b"]
spec:
minMember: 5
minResources:
cpu: "5"
memory: "2048Mi"
scheduleTimeoutSeconds: 600

It should be noted that, if use Gang feature by CRD way, user should let high level operator maintain Gang CRD life circle like handling update/create/delete events. Also, from a Scheduler perspective, scheduler should handle receive-order-issue's between Gang CRD and pod. For example, if pods arrive to scheduler before Gang CRD, we have to build a fake Gang data structure temporarily to collect all related pods, and need to suspend the scheduling of pods until parse the configuration from real Gang CRD.

Annotation way​

gang.scheduling.koordinator.sh/name           
gang.scheduling.koordinator.sh/min-available

The upper definitions are indispensable. We are compatible with pod-group.scheduling.sigs.k8s.io, pod-group.scheduling.sigs.k8s.io/name and pod-group.scheduling.sigs.k8s.io/min-available in community. We also support new definitions to declare Gang's name and minimum number.

Also, we introduce some optional definitions as below, most are mentioned above:

gang.scheduling.koordinator.sh/waiting-time
gang.scheduling.koordinator.sh/total-number
gang.scheduling.koordinator.sh/mode
gang.scheduling.koordinator.sh/groups
  • gang.scheduling.koordinator.sh/waiting-time represents max wait time since first pod comes to permit stage. Default is a global config.

  • gang.scheduling.koordinator.sh/total-number helps to calculate Gang scheduling cycle in strict mode, you can find more detail in Data-Structure chapter. Default equals to gang.scheduling.koordinator.sh/min-available.

  • gang.scheduling.koordinator.sh/mode determines Strict or NonStrict. Default is Strict.

  • gang.scheduling.koordinator.sh/groups describes GangGroups. Default is empty, which means don't need to form a GangGroup with others.

It should be noted that, the annotation mode's parameter will overwrite CRD's mode if both exist. And gangGroup should be announced with " gangNamespace" + "/" + "gangName "

Example​

When user apply a basic gang, the example is as follows:

metadata:
annotations:
gang.scheduling.koordinator.sh/name: gang-a
gang.scheduling.koordinator.sh/min-available: 5

Let's assume a job has two roles: A and B, each role has several pods. PodA belongs to roleA, podB belongs to roleB. roleA and roleB belongs to one GangGroup, the example is as follows:

metadata:
annotations:
gang.scheduling.koordinator.sh/name: gang-a
gang.scheduling.koordinator.sh/waiting-time: 3600s
gang.scheduling.koordinator.sh/min-available: 5
gang.scheduling.koordinator.sh/total-number: 5
gang.scheduling.koordinator.sh/mode: Strict
gang.scheduling.koordinator.sh/groups: ["namespaceA/gang-a", "namespaceB/gang-b"]
metadata:
annotations:
gang.scheduling.koordinator.sh/name: gang-b
gang.scheduling.koordinator.sh/waiting-time: 3600s
gang.scheduling.koordinator.sh/min-available: 5
gang.scheduling.koordinator.sh/total-number: 5
gang.scheduling.koordinator.sh/mode: Strict
gang.scheduling.koordinator.sh/groups: ["namespaceA/gang-a", "namespaceB/gang-b"]

Assuming a job has two roles: A and B, each role has several pods. podA belongs to roleA, podB belongs to roleB. roleA and roleB belongs to different GangGroup, the example as follows:

metadata:
annotations:
gang.scheduling.koordinator.sh/name: gang-a
gang.scheduling.koordinator.sh/waiting-time: 3600s
gang.scheduling.koordinator.sh/min-available: 5
gang.scheduling.koordinator.sh/total-number: 5
gang.scheduling.koordinator.sh/mode: Strict
gang.scheduling.koordinator.sh/groups: ""
metadata:
annotations:
gang.scheduling.koordinator.sh/name: gang-b
gang.scheduling.koordinator.sh/waiting-time: 3600s
gang.scheduling.koordinator.sh/min-available: 5
gang.scheduling.koordinator.sh/total-number: 5
gang.scheduling.koordinator.sh/mode: Strict
gang.scheduling.koordinator.sh/groups: ""

Implementation Details​

QueueSortPlugin​

We design an independent plugin to implement the QueueSort extension point separately, so that we can integrate queue sort logic of all plugins, and register them at one time.

In this proposal, we implement the Less function to gather pods belong to same Gang. The specific queuing rule is:

  1. Firstly, compare the priorities of the two pods, the higher priority is at the front of the queue.

  2. Secondly, compare creationTimestamp of two pods, if pod belongs to a Gang, then we compare creationTimestamp of the Gang, the one created first will be at the front of the queue.

  3. Finally, compare pod's namespace, if pod belongs to a Gang, then we compare Gang name.

type QueueSortPlugin interface{
QueueSort(*QueuedPodInfo, *QueuedPodInfo) bool
}

GangSchedulingPlugin​

Data-Structure​
Gang​
type Gang struct {
Name string
WaitTime time.Duration
Mode string //Strict or NonStrict
GangGroup []string
MinRequiredNumber int
TotalChildrenNum int
Children map[string]*PodInfo
BoundChildren map[string]*PodInfo
WaitingForBindChildren map[string]*PodInfo
ResourceSatisfied bool
ScheduleCycle int
ScheduleCycleValid bool
ChildrenScheduleRoundMap map[string]int
}

We design the Gang to record Gang status in scheduler memory. We can get the children pods from "Children" field, and the BoundChildren, WaitingForBindChildren store the pods binding status, which is used to check if the pods can pass permit stage.

Once Permit stage passed, we will set ResourceSatisfied=true, as mentioned above in After Gang chapter, this variable is used for judging whether gang has been satisfied. when handle failover case, if any pod in Gang has been bound, we set ResourceSatisfied=true.

We especially explain scheduleCycle and childrenScheduleRoundMap field. These fields control Gang's scheduling cycle. For example, at the beginning, scheduleCycle is 1, and each pod's cycle in childrenScheduleRoundMap is 0. When each pod comes to PreFilter, we will check if the pod's value in childrenScheduleRoundMap is smaller than Gang's scheduleCycle, If result is positive, we set the pod's cycle in childrenScheduleRoundMap equal with scheduleCycle and pass the check. If result is negative, means the pod has been scheduled in this cycle, so we should reject it. With totalChildrenNum's help, when the last pod comes to make all childrenScheduleRoundMap's values equal to scheduleCycle, Gang's scheduleCycle will be added by 1, which means a new schedule cycle.

We continue to explain scheduleCycleValid field, during the scheduling, When a pod failed at Filter stage, we will set ScheduleCycleValid to false in PostFilter stage, which means any pod in this Gang shouldn't be scheduled until it is set to "true", and the remaining pods should be rejected in PreFilter stage. Only When scheduleCycle added by 1, we will reset the scheduleCycleValid to true.

It should be emphasized that scheduleCycle\scheduleCycleValid\childrenScheduleRoundMap only work in Strict.

GangPlugin​

this is the framework of the Plugin,we cache the Gang info above in the gangCache.

type GangPlugin struct {
frameworkHandler framework.Handle
gangClient gangClient.Interface
podLister listerv1.PodLister
snapshotSharedLister framework.SharedLister
gangCache map[string]*Gang
}

during the whole kubernetes shceduling process,we only need to realize our logic into four extention points as below:

var(
_ framework.PreFilterPlugin = &GangScheduling{}
_ framework.PostFilterPlugin = &GangScheduling{}
_ framework.PermitPlugin = &GangScheduling{}
_ framework.ReservePlugin = &Coscheduling{}
)
type GangScheduling interface{
ActiveGang(pod *corev1.Pod, state *framework.CycleState)
PreFilter(context.Context, *corev1.Pod) error
PostFilter(ctx context.Context, state *CycleState, pod *v1.Pod, filteredNodeStatusMap NodeToStatusMap) (*PostFilterResult, *Status)
Permit(context.Context, *corev1.Pod) Status
Unreserve(ctx context.Context, state *framework.CycleState, pod *v1.Pod, nodeName string)
}
PreFilter​

if NonStrict, we only do step1 and step2:

  • Check whether childes in Gang has met the requirements of minimum number under each Gang, and reject the pod if negative.

  • Check whether the Gang has been timeout(check the pod's annotation,later introduced at Permit section), and reject the pod if positive.

  • Check whether the Gang has met the scheduleCycleValid check, and reject the pod if negative.

  • Try update scheduleCycle, scheduleCycleValid, childrenScheduleRoundMap as mentioned above.

PostFilter​

At this point means the pod didn't pass the Filter Plugin, we should:

  • If Strict, we will set scheduleCycleValid to false and release all assumed pods.

  • If NonStrict, we will do nothing.

Permit​

Any pod passes Filter stage will come to this stage. Scheduler will calculate all Gangs in GangGroup whether the current number of assumed-pods in each Gang meets the Gang's minimum requirement.

  • If Gang don't meet the bind-condition, we will give the pod a "Wait" Status with a timeout duration, and the bind goroutine will keep waiting until the wait is timeout or passed. Then we will run the ActiveGang method, it can put all the pods belong to the Gang which in schedulableQueue or backoffQueue back to activeQueue, so that the pod of Gang can be continuously scheduled as much as possible.

It should be noted that, in community, scheduler limit maximum timeout value under 15 min, we may need to hook RunPermitPlugins to enlarge the timeout when 15 minutes is not enough. Now we record as a known-issue.

  • If Gang meet the bind-condition, we will give every waiting pod a "Success" status, which will let the bind goroutine of each pod leave the waiting status and continue to run. Also, as mentioned above, we will set Gang's ResourceSatisfied to true.
Un-reserve​

Both permit stage is timeout and binding failed will lead the pod to un-reserve stage, we can distinguish from Gang's "ResourceSatisfied" field, if the field is true means binding failed, else means the Gang is timeout.

  • When permit stage is timeout, we will give an annotation like gang.scheduling.koordinator.sh/timeout=true to all the pods belong to the Gang and will release the resource of all the assumed pods. The Gang will not be scheduled anymore, user should manually handle the timeout event.

  • When binding failed, as mentioned above, the collection of Gang's resource is over, we will do nothing except roll back the failed pod resource.

Init​

We will register pod's event handler to watch pod event for updating Gang.

Unsolved Problems​

Alternatives​

User can choose use Gang by Strict and NonStrict case by case.