Wei Huang
Ex-lead of Kubernetes sig-scheduling
Menlo Park, California, United States
Actions
Wei Huang is a Software Engineer at Apple, focusing on Kube scheduling and control plane. He has served as a co-chair of Kubernetes SIG-Scheduling for years. He is also the founder of two Kubernetes sub-projects, scheduler-plugins, and kwok.
Links
Exploring tenant-centric strategies to simplify multi-cluster and multi-cloud complexities
Kubernetes is widely recognized as a platform for building platforms, but even with modern platform engineering techniques, managing an end-to-end release and deployment lifecycle remains challenging.
This talk will first analyze key pain points in platform engineering and propose a new perspective on infrastructure code by separating different personas’ views. We advocate for a tenant-centric API that prioritizes user experience—minimizing input, reducing learning curves, and abstracting cloud provider details to decouple desired resources from underlying specs.
Next, we’ll introduce a design for fanning out tenant resource claims, enhancing flexibility and extensibility through a code generation component and a Kubernetes-like labeling system. Finally, we’ll cover the glue that binds everything together: Pkl for templating and validation, Prow for GitHub event-driven automation, and Crossplane + ArgoCD as the claim realization engine.
Topology-Aware Scheduling for AI Training & Inference with Kueue
Kueue, the Kubernetes-native workload orchestrator, helps run AI workloads at scale with multi-tenant quota management and advanced scheduling.
In this session, we show through a case study at the Meta Superintelligence Lab how Kueue provided the right quota and scheduling layer for large-scale AI training and inference across complex, heterogeneous GPU topologies, enabling a shift from an in-house platform to an open-source stack on Kubernetes.
We then dive into Kueue’s multi-layer Topology-Aware Scheduling (TAS) for modern clusters with hierarchical network fabrics. We cover the main concepts and algorithms behind Hierarchical TAS, and show why multi-layer TAS was critical for this transition.
We close with other cutting-edge Kueue capabilities, including Workload-Aware Scheduler integration, elastic workloads, and multi-cluster job scheduling with MultiKueue.
SIG-Scheduling Intro & Deep Dive
Kube-scheduler is a critical component to Kubernetes, responsible for placing the pod to the most suitable node. But how it works, can we customize it for advanced usage, what’s the best practice in large clusters. To answer these progressive questions, we’ll divide this session into two parts. If you’re a newbie to kube-scheduler, you may interest with the Intro part, if you’re a senior one, you can join our Deep Dive.
What’s more, we’ll share with you some ongoing works within the SIG, including the latest progress with the sub-projects.
A Comparative Analysis of Kueue, Volcano, and YuniKorn
Choosing the best solution for running batch workloads on Kubernetes can be frustrating. Kueue, Volcano, and Apache YuniKorn were designed to address similar challenges but differ in how they tackle them. Deciding which is most suitable for a specific use case is often confusing.
Batch workloads like big data, data engineering, HPC, AI, and machine learning share common requirements, especially around batch-scheduling. Managing resource sharing and isolation between tenants while balancing utilization and meeting SLAs presents a significant challenge on Kubernetes.
This session dives into three community-driven solutions: Kueue, Volcano, and Apache YuniKorn. We’ll explore their features, use-case suitability, and design trade-offs, providing a comprehensive comparison. Attendees will leave with the insights needed to answer a crucial question: which solution best addresses the batch-scheduling needs of my workloads?
SIG-Scheduling Intro & Deep Dive
Kube-scheduler is a critical component to Kubernetes, responsible for placing the pod to the most suitable node. But how it works, can we customize it for advanced usage, what’s the best practice in large clusters. To answer these progressive questions, we’ll divide this session into two parts. If you’re a newbie to kube-scheduler, you may interest with the Intro part, if you’re a senior one, you can join our Deep Dive.
What’s more, we’ll share with you some ongoing works within the SIG, including the latest progress with the sub-projects.
Revolutionizing Kube Scalability Testing with KWOK
Scalability testing in Kubernetes has long been considered a demanding and enigmatic process. In this presentation, we aim to dispel the obscurity surrounding this subject, beginning with a comprehensive analysis of existing solutions such as Kubemark. We will explore potential areas for improvement: for instance, transforming from N hollow nodes to a single central controller while maintaining a memory footprint under 100MB, thus making it possible to conduct a 5k node/50k pod test on a laptop.
We will then illustrate the utilization of KWOK for YuniKorn's scalability testing. Given KWOK's low resource demands, we will demonstrate its seamless integration into popular CI pipelines such as Github Actions. This allows early detection of performance degradation within the development workflow. Our discourse aims to elucidate a path toward more efficient and effective Kubernetes scalability testing and set the stage for innovative developments in the field.
Wei Huang
Ex-lead of Kubernetes sig-scheduling
Menlo Park, California, United States
Links
Actions
Please note that Sessionize is not responsible for the accuracy or validity of the data provided by speakers. If you suspect this profile to be fake or spam, please let us know.
Jump to top