Stenobird

Episode

HPC Workload Scheduling, with Ricardo Rocha

Podcast
Kubernetes Podcast from Google
Published
Jul 9, 2025
Duration seconds
2578
Processing state
processed
Canonical source
https://e780d51f-f115-44a6-8252-aed9216bb521.libsyn.com/hpc-workload-scheduling-with-ricardo-rocha
Audio
https://traffic.libsyn.com/secure/e780d51f-f115-44a6-8252-aed9216bb521/KPOD255.mp3?dest-id=3486674
JSON
/v1/public/podcasts/kubernetes-podcast-from-google/episodes/hpc-workload-scheduling-with-ricardo-rocha
Markdown
/podcast/kubernetes-podcast-from-google/hpc-workload-scheduling-with-ricardo-rocha.md

Actions

  • POST https://stenobird.com/v1/public/podcasts/kubernetes-podcast-from-google/episodes/hpc-workload-scheduling-with-ricardo-rocha/transcription-requests
    Idempotently request low-priority transcript generation for this episode.
  • GET https://stenobird.com/podcast/kubernetes-podcast-from-google/hpc-workload-scheduling-with-ricardo-rocha.md
    Read the agent-friendly Markdown representation of this episode resource.

Summary

Ricardo Rocha from CERN explains how to bridge the gap between traditional High-Performance Computing (HPC) and cloud-native Kubernetes environments. The discussion focuses on using specialized schedulers like Kueue to manage expensive, high-demand resources like GPUs without reinventing the Kubernetes core.

Topics

  • Kubernetes
  • HPC
  • CERN
  • GPU Scheduling
  • Kueue
  • Cloud Native
  • CNCF
  • Batch Workloads
  • Infrastructure Engineering

Highlights

  • Main idea: Kubernetes is evolving from a stateless web-app orchestrator to a platform capable of handling complex, resource-intensive scientific workloads
  • Practical takeaway: Use 'out-of-tree' controllers like Kueue to implement specialized logic (like pod suspension) while leveraging native Kubernetes primitives to avoid maintenance debt
  • Failure mode: Building parallel execution systems that bypass Kubernetes semantics leads to massive technical debt as you must manually reimplement every core upstream improvement
  • Main idea: The rise of expensive, pre-committed GPU instances is shifting cloud computing from an 'on-demand' model back toward a 'reservation' model similar to on-premises hardware
  • Practical takeaway: Focus on extending Kubernetes via the Gateway API and specialized controllers rather than fighting the existing system's architecture

Chapters

  1. 1:00 Node Feature Discovery and Gemini CLI: An introduction to NFD for hardware-aware scheduling and the new Gemini CLI for interacting with AI from the terminal.
  2. 4:05 CERN's Resource Challenges: Discussing the necessity of optimizing fixed budgets and managing massive datasets in scientific computing.
  3. 7:20 Low-Level Optimization in HPC: The shift toward needing CPU pinning, NUMA awareness, and high-efficiency node usage for research workloads.
  4. 10:25 The Evolution of Kubernetes for Research: Why it took time for scientific and research-oriented voices to integrate into the Kubernetes ecosystem.
  5. 13:50 Managing Expensive GPU Resources: The difficulty of managing heterogeneous hardware and preventing over-provisioning of high-cost accelerators.
  6. 17:00 The Role of Specialized Schedulers: A look at the history of grid computing and how projects like Volcano and UniKorn provide necessary extensions.
  7. 20:05 The CNCF Ecosystem and Maturity: How the CNCF sandbox and incubation process supports the growth of specialized cloud-native projects.
  8. 23:20 Future of Batch Workloads: Speculating on multi-cluster, multi-region management and the move toward resource reservation models.