Computational Fluid Dynamics (CFD) Analysis With Kubernetes, Kubeflow, And OpenFOAM
summary
Kubeflow’s MPI operator runs OpenFOAM CFD jobs on an OpenShift Kubernetes cluster by creating a launcher pod that SSH‑bootstraps work across MPI operator launcher with SSH bootstrap into many workers. The experiment uses a configMap‑injected decomposer script and Open Data Foundation RWX storage, launches 256 worker pods / 512 CPU ranks, and runs on cloud VMs. The run completes but takes ~5 hours and shows poor throughput from overslicing and SSH race conditions; PMIx and GPU acceleration are suggested next steps.
AI-written summary · about
chapters (16)
- 0:00Introduction
- 1:03Experiment building blocks
- 3:32OpenFOAM MPI and Kubernetes basics
- 5:56MPI operator workflow
- 7:19Dam Break CFD example
- 9:10Performance tuning and AWS
- 10:40Future directions
- 12:45Q&A
- 13:29Next steps and automation
- 15:14Red Hat’s HPC on Kubernetes
- 17:20Comparison with Morin’s workflow
- 21:33Scheduling and resource contention
- 25:48CubeFlow MPI operator
- 29:29OpenFOAM container images
- 30:55High-performance storage and Infiniband
- 33:40MPI bootstrapping race conditions
related talks
40mKubeflow: ML on OpenShift
29mKubeflow: TensorFlow on Kubernetes
51mMachine Learning using Kubeflow and Kubernetes
23mEfficient Integration of Kubeflow Tools: An ISS Data Case Study on Anomaly Detection
34mKubeflow: Multi-Tenant, Self-Serve, Accelerated Platform for Practitioners
14mKubeflow: Where ML Meets Kubernetes
31mKubeflow - Data Science on Steroids
15mLightning Talk: Streamlining MLOps Pipeline With Kubeflow on Arm64 Locally