Explore Kueue, a Kubernetes-native job queueing system, in this informative conference talk. Dive into the challenges of managing batch workloads in Kubernetes and learn how Kueue addresses limitations in pod-centric scheduling. Discover the concept of job queueing and its benefits for resource management, especially for GPU-intensive tasks. Understand Kueue's design principles, API structure, and integration with existing Kubernetes components. Gain insights into resource allocation across multiple namespaces and clusters, borrowing cohorts, and autoscaling capabilities. Learn about the workload API, custom workload handling, and capacity management. Explore how Kueue interacts with index jobs and enhances overall cluster efficiency for batch processing tasks.
Kueue: Kubernetes-Native Job Queueing for Batch Workloads
CNCF [Cloud Native Computing Foundation] via YouTube
Overview
Syllabus
Introduction
Problem description
What is a job
Batch workloads
Flexible on multiple dimensions
The problem
Microbatch jobs
What is Job Queueing
What we want to solve
Why propose a new controller
The concept of a queue
Design principle
Job API
Resource Model
Multiple namespaces
Multiple cluster queues
Borrowing cohorts
Queue API
Cluster Queue API
Dakota
Borrowing Cohort
How it works
suspend
queuing controller
existing controller
handling custom workloads
workload API
workload controller
autoscaling
why
integration
capacity requests
pod management
release
questions
how does the queue interact with index jobs
Taught by
CNCF [Cloud Native Computing Foundation]
Related Courses
-
Building a Batch System for the Cloud with Kueue
-
Sailing Ray Workloads with KubeRay and Kueue in Kubernetes
-
Batch Systems in Production with Kueue - Multi-Tenancy and Fungibility
-
Kueue Controller and Kubernetes Batch Working Group
-
Advanced Resource Management for Running AI/ML Workloads with Kueue
-
Scaling Batch, Big Data, and AI Workloads Beyond the Kubernetes Scheduler
