Preemption

In a preemption, the following terms are relevant:

• Preemptees: The preempted Workloads.
• Target ClusterQueues: The ClusterQueues to which the preemptees belong.
• Preemptor: The Workload being accommodated.
• Preempting ClusterQueue: The ClusterQueue to which the preemptor belongs.

Reasons for preemption

A Workload can preempt one or more Workloads if it is admitted in a ClusterQueue with preemption enabled and any of the following events happen:

• The preemptee belongs to the same ClusterQueue as the preemptor and the preemptee has a lower priority.
• The preemptee belongs to the same cohort as the preemptor and the preemptee’s ClusterQueue has a usage above the nominal quota for at least one resource that the preemptee and preemptor require.

The configured settings for preemption in the Kueue Configuration and in the ClusterQueue can limit whether a Workload can preempt others, in addition to the criteria above.

When preempting a Workload, Kueue adds entries in the .status.conditions field of the preempted Workload that is similar to the following:

status:
  conditions:
  - lastTransitionTime: "2024-05-31T18:42:33Z"
    message: 'Preempted to accommodate a workload (UID: 5515f7da-d2ea-4851-9e9c-6b8b3333734d)
      in the ClusterQueue'
    observedGeneration: 1
    reason: Preempted
    status: "True"
    type: Evicted
  - lastTransitionTime: "2024-05-31T18:42:33Z"
    message: 'Preempted to accommodate a workload (UID: 5515f7da-d2ea-4851-9e9c-6b8b3333734d)
      in the ClusterQueue'
    reason: InClusterQueue
    status: "True"
    type: Preempted

The Evicted condition indicates that the Workload was evicted with a reason Preempted, whereas the Preempted condition gives more details about the preemption reason.

Preemption algorithms

Kueue offers two preemption algorithms. The main difference between them is the criteria to allow preemptions from a ClusterQueue to others in the Cohort, when the usage of the preempting ClusterQueue is already above the nominal quota. The algorithms are:

• Classic Preemption: Preemption in the cohort only happens when the usage of the preempting ClusterQueue will be under the nominal quota after the ongoing admission process, or when all the candidates for preemption belong to the same ClusterQueue as the preempting Workload. In other words, ClusterQueues can only borrow quota from others in the cohort if they do not preempt admitted Workloads from other ClusterQueues. ClusterQueues in a cohort borrow resources in a first-come first-served fashion. This algorithm is the most lightweight of the two.
• Fair sharing: ClusterQueues with pending Workloads can preempt other Workloads in their cohort until the preempting ClusterQueue obtains an equal or weighted share of the borrowable resources. The borrowable resources are the unused nominal quota of all the ClusterQueues in the cohort.

Classic Preemption

An incoming Workload, which does not fit within the unused quota, is eligible to issue preemptions when one of the following is true:

• the requests of the Workload are below the flavor’s nominal quota, or
• borrowWithinCohort is enabled.

Candidates

The list of preemption candidates is compiled from Workloads within the Cluster Queue satisfying the withinClusterQueue policy, and Workloads within the cohort which satisfy the reclaimWithinCohort policy.

The list of candidates is sorted based on the following preference checks for tie-breaking:

• Workloads from borrowing queues in the cohort
• Workloads with the lowest priority
• Workloads which got admitted the most recently.

Targets

The Classic Preemption algorithm qualifies the candidates as preemption targets using the heuristics below:

1. If all candidates belong to the target queue, then Kueue greedily qualifies candidates until the preemptor Workload can fit, allowing the usage of the ClusterQueue to be above the nominal quota, up to the borrowingLimit. This is referred as “borrowing” in the points below.

2. If borrowWithinCohort is enabled, then Kueue greedily qualifies candidates (respecting the borrowWithinCohort.maxPriorityThreshold threshold), until the preemptor Workload can fit, allowing for borrowing.

3. If the current usage of the target queue is below nominal quota, then Kueue greedily qualifies the candidates, until the preemptor Workload can fit, disallowing for borrowing.

4. If the Workload didn’t fit by using the previous heuristics, Kueue greedily qualifies only the candidates which belong to the preempting Cluster Queue, until the preemptor Workload can fit, allowing for borrowing.

The last step of the algorithm is to minimize the set of targets. For this purpose, Kueue greedily traverses the list of initial targets in reverse and removes a Workload from the list of targets if the preemptor Workload still can be admitted when accounting back the quota usage of the target Workload.

Fair Sharing

To enable fair sharing, use a Kueue Configuration similar to the following:

apiVersion: config.kueue.x-k8s.io/v1beta1
kind: Configuration
fairSharing:
  enable: true
  preemptionStrategies: [LessThanOrEqualToFinalShare, LessThanInitialShare]

The attributes in this Kueue Configuration are described in the following sections.

ClusterQueue share value

When you enable fair sharing, Kueue assigns a numeric share value to each ClusterQueue to summarize the usage of borrowed resources in a ClusterQueue, in comparisson to others in the same cohort. The share value is weighted by the .spec.fairSharing.weight defined in a ClusterQueue.

During admission, Kueue prefers to admit Workloads from ClusterQueues that have the lowest share value first. During preemption, Kueue prefers to preempt Workloads from ClusterQueues that have the highest share value first.

You can obtain the share value of a ClusterQueue in the .status.fairSharing.weightedShare field or querying the kueue_cluster_queue_weighted_share metric.

Preemption strategies

The preemptionStrategies field in the Kueue Configuration indicates which constraints should a preemption satisfy, with regards to the share values of the target and preempting ClusterQueues, before and after preempting a particular Workload.

Different preemptionStrategies can lead to less or more preemptions under specific scenarios. These are the factors you should consider when configuring preemptionStrategies:

• Tolerance to disruptions, in particular when single Workloads use a significant amount of the borrowable resources.
• Speed of convergence, in other words, how important is it to reach a steady fair state as soon as possible.
• Overall utilization, because certain strategies might reduce the utilization of the cluster in the pursue of fairness.

When you define multiple preemptionStrategies, the preemption algorithm will only use the next strategy in the list if there aren’t any more Workloads that are candidates for preemption that satisfy the current strategy and the preemptor still doesn’t fit.

The values you can put in the preemptionStrategies list are:

• LessThanOrEqualToFinalShare: Only preempt a Workload if the share of the preempting ClusterQueue with the preemptor Workload is less than or equal to the share of the target ClusterQueue without the preempted Workload. This strategy might favor preemption of smaller workloads in the target ClusterQueue, regardless of priority or start time, in an effort to keep the share of the ClusterQueue as high as possible.
• LessThanInitialShare: Only preempt a Workload if the share of the preempting ClusterQueue with the preemptor Workload is strictly less than the share of the target ClusterQueue. Note that this strategy doesn’t depend on the share usage of the Workload being preempted. As a result, the strategy chooses to first preempt workloads with the lowest priority and newest start time within the target ClusterQueue. The default strategy is [LessThanOrEqualToFinalShare, LessThanInitialShare]

Algorithm overview

The initial step of the algorithm is to identify the Workloads that are candidate for preemption, with the same criteria and ordering as the classic preemption, and grouped by ClusterQueue.

Next, the above candidates are qualified as preemption targets, following an algorithm that can be summarized as follows:

FindFairPreemptionTargets(X ClusterQueue, W Workload)
  For each preemption strategy:
    While W does not fit and there are workloads that are preemption candidates:
      Find the ClusterQueue Y with the highest share value.
      For each admitted Workload U in ClusterQueue Y:
        If Workload U satisfies the preemption strategy:
          Add workload U to the list of targets
  In the reverse order of the list of targets:
    Attempt to remove a Workload from the targets, while W still fits.