fix(enqueue): capacity plugin job enqueueable

[hajnalmt]

What type of PR is this?

/kind bug
/kind enhancement

What this PR does / why we need it:

The PR contains some enhancements and a bugfix regarding the capacity plugin's enqueue process.
Currently in the capacity plugin we rely purely on realCapability to decide wheter a job shall enqueue or not.

This is a little bit problematic, as there are cases when we want to enqueue below deserved or guarantee directly as
zjj2wry points out here: #4817 (comment)
For example, when we have a queue structure where:

# Cluster: 100 CPU
parent (guarantee = 100 CPU, capability = 100 CPU)
├── leaf-1:
│   - guarantee = 60 CPU
│   - deserved = 60 CPU
│   - allocated = 100 CPU (overused by 40 CPU)
│
└── leaf-2:
    - guarantee = 40 CPU
    - deserved = 40 CPU
    - allocated = 0 CPU
    - wants 40 CPU

The leaf-2 queue cannot enqueue, although having guaranteed resources on the cluster.
There can be use cases where we don't guarantee resources on the cluster to any queues, rather than we are slicing the cluster on deserved values, for example if we omit guarantee in the above example:

# Cluster: 100 CPU
parent (guarantee = 100 CPU, capability = 100 CPU)
├── leaf-1:
│   - deserved = 60 CPU
│   - allocated = 100 CPU (overused by 40 CPU)
│
└── leaf-2:
    - deserved = 40 CPU
    - allocated = 0 CPU
    - wants 40 CPU

leaf-2 in this case will not able to enqueue too due to maxed out cluster capacity, although it would deserve 40
For this use case (when no guarantee is specified) we should enqueue too I think.

The comparision uses LessEqualPartlyWithDimensionZeroFiltered so if there exists any dimension that is below or equal
in deserved/guarantee we still enqueue. This means in theory that there can be dimensions where the job's minResources will increase the allocated+ inqueued queue resources above the queue capability value, which is fine I think as this is the realm of allocate rather than enqueue, so the principle that we don't let resources to be created in a queue above the capability is not harmed.

Additionally, these examples are for hierarchical queues, but the causes and the induction is true for the non-hierarchical cases too.

Minor enhancement in the enqueue action logging and the capacity plugins tests has been added to the PR.

Which issue(s) this PR fixes:

Fixes #4817

Special notes for your reviewer:

Does this PR introduce a user-facing change?

fix(enqueue): capacity plugin job enqueueable
A bug has been fixed in the capacity plugin's enqueue logic, which prohibited legitimate jobs to move from `Pending` to `Inqueue`.
Additionally we added a new logic which will let the jobs to enqueue below `deserved`, when no `guarantee` value is provided in any of the queues across cluster.

[volcano-sh-bot]

@hajnalmt: The label(s) kind/enhancement cannot be applied, because the repository doesn't have them.

Details

In response to this:

What type of PR is this?

/kind bug
/kind enhancement

What this PR does / why we need it:

The PR contains some enhancements and a bugfix regarding the capacity plugins enqueue process.
Currently in the capacity plugin we are relying on realCapability to decide wheter a job shall enqueued or not.

This is a little bit problematic, as there are cases when we want to enqueue below deserved or guarantee directly as
zjj2wry points out here: #4817 (comment)
For example, when we have a queue structure where:

# Cluster: 100 CPU
parent (guarantee = 100 CPU, capability = 100 CPU)
├── leaf-1:
│   - guarantee = 60 CPU
│   - deserved = 60 CPU
│   - allocated = 100 CPU (overused by 40 CPU)
│
└── leaf-2:
   - guarantee = 40 CPU
   - deserved = 40 CPU
   - allocated = 0 CPU
   - wants 40 CPU

The leaf-2 queue cannot enqueue, although having guaranteed resources on the cluster.
There can be use cases where we don't guarantee resources on the cluster to any queues, rather than we rely only on deserved values, if we omit guarantee in the above example:

# Cluster: 100 CPU
parent (guarantee = 100 CPU, capability = 100 CPU)
├── leaf-1:
│   - deserved = 60 CPU
│   - allocated = 100 CPU (overused by 40 CPU)
│
└── leaf-2:
   - deserved = 40 CPU
   - allocated = 0 CPU
   - wants 40 CPU

leaf-2 in this case will not able to enqueue too due to maxed out cluster capacity, although it would deserve 40
For this use case (when no guarantee is specified) we should enqueue too I think.

The comparision uses LessEqualPartlyWithDimensionZeroFiltered so if there exists any dimension that is below or equal
in deserved/guarantee we still enqueue. This means in theory that there can be dimensions where the job's minResources will increase the inqueued queue resources above the queue capability value, which is fine I think as this is the realm of allocate rather than enqueue, so the principle that we don't let resources to be created in a queue above the capability is not harmed.

Additionally, these examples are for hierarchical queues, but the causes and the induction is true for the non-hierarchical cases too.

Which issue(s) this PR fixes:

Fixes #4817

Special notes for your reviewer:

Does this PR introduce a user-facing change?

fix(enqueue): capacity plugin job enqueueable
A bug has been fixed in the capacity plugin's enqueue logic, which prohibited legitimate jobs to move from `Pending` to `Inqueue`. Additionally we added a logic which will let the jobs to enqueue, when no `guarantee` value is provided in any of the queues across cluster.

Instructions for interacting with me using PR comments are available here. If you have questions or suggestions related to my behavior, please file an issue against the kubernetes/test-infra repository.

[gemini-code-assist]

Summary of Changes

Hello @hajnalmt, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request significantly improves the job enqueueing process within the capacity plugin by resolving a bug that prevented legitimate jobs from being enqueued under certain resource allocation conditions. It introduces more flexible resource evaluation, allowing jobs to progress when their requested resources fall within a queue's guaranteed or deserved limits, even if the overall capacity is temporarily oversubscribed. This enhancement ensures more efficient utilization of cluster resources and a smoother workflow for jobs in various queue configurations.

Highlights

• Enhanced Job Enqueueing Logic: The capacity plugin's job enqueueing mechanism has been updated to allow jobs to be enqueued based on "guarantee" or "deserved" resources, even if the "realCapability" is currently exceeded. This addresses scenarios where jobs were incorrectly blocked from moving from "Pending" to "Inqueue".
• Flexible Resource Evaluation: The "jobEnqueueable" function now uses "LessEqualPartlyWithDimensionZeroFiltered" to evaluate resource availability, meaning a job can be enqueued if it fits within "guarantee" or "deserved" for at least one resource dimension, preventing unnecessary blocking.
• Improved Hierarchical Queue Handling: The enqueueing logic has been refined for hierarchical queues, ensuring that checks against "guarantee", "deserved", and "realCapability" are correctly applied across the queue hierarchy.
• Comprehensive Test Coverage: New test cases have been added to validate the updated enqueueing behavior, specifically covering scenarios involving "guarantee" and "deserved" resources, both when enqueuing is expected and when it should be prevented due to cluster saturation.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description
Code Review	/gemini review	Performs a code review for the current pull request in its current state.
Pull Request Summary	/gemini summary	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help	/gemini help	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

Footnotes

1. Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

[gemini-code-assist]

Code Review

This pull request introduces a valuable fix and enhancement to the capacity plugin's enqueue logic, allowing jobs to be enqueued based on guarantee or deserved resources, which resolves issues in hierarchical queue scenarios. The changes are well-supported by a comprehensive set of new test cases that cover the described scenarios. My review includes a suggestion to refactor a part of the new logic to improve maintainability by reducing code duplication.

[zjj2wry]

LGTM, the only concern is excessive enqueueing when GPU is insufficient but CPU is sufficient. However, I can configure child queues with GPU guarantee only

[volcano-sh-bot]

[APPROVALNOTIFIER] This PR is NOT APPROVED

This pull-request has been approved by: zjj2wry
Once this PR has been reviewed and has the lgtm label, please assign jessestutler for approval. For more information see the Kubernetes Code Review Process.

The full list of commands accepted by this bot can be found here.

Details

Needs approval from an approver in each of these files:

• pkg/scheduler/OWNERS

Approvers can indicate their approval by writing /approve in a comment
Approvers can cancel approval by writing /approve cancel in a comment

[JesseStutler]

I'm wondering that in this case:

# Cluster: 100 CPU
parent (guarantee = 100 CPU, capability = 100 CPU)
├── leaf-1:
│   - guarantee = 60 CPU
│   - deserved = 60 CPU
│   - allocated = 100 CPU (overused by 40 CPU)
│
└── leaf-2:
    - guarantee = 40 CPU
    - deserved = 40 CPU
    - allocated = 0 CPU
    - wants 40 CPU

Why leaf-1 can allocate 100CPU? Because leaf-2 has 40CPU guarantee resources, leaf-1's realCapability should be 60CPU, why leaf-1 can allocate 100CPU?

BTW, I think the case is very weird, we prefer to comply guarantee < deserved < capability, don't know we have such a real scenario totalGuarantee == totalResource

[zjj2wry]

I'm wondering that in this case:

# Cluster: 100 CPU
parent (guarantee = 100 CPU, capability = 100 CPU)
├── leaf-1:
│   - guarantee = 60 CPU
│   - deserved = 60 CPU
│   - allocated = 100 CPU (overused by 40 CPU)
│
└── leaf-2:
    - guarantee = 40 CPU
    - deserved = 40 CPU
    - allocated = 0 CPU
    - wants 40 CPU

Why leaf-1 can allocate 100CPU? Because leaf-2 has 40CPU guarantee resources, leaf-1's realCapability should be 60CPU, why leaf-1 can allocate 100CPU?

BTW, I think the case is very weird, we prefer to comply guarantee < deserved < capability, don't know we have such a real scenario totalGuarantee == totalResource

The scenario where leaf-2 is created after leaf-1 has already allocated resources. when cluster resources are insufficient and new nodes are added through scaling, existing pending tasks may immediately consume all the newly added resources, preventing new queues from scheduling pods even if they have guaranteed resources.

[zjj2wry]

the job will always enqueue like below. i think we should checks all dimensions of resources when determining if a job can be enqueue and preempt.

guarantee: CPU=100, Memory=200Gi, GPU=8
used: CPU=0, Memory=0, GPU=8 (GPU exhausted)
request: CPU=10, Memory=20Gi, GPU=2