Lightning Talk: Fluence: Approaching a Converged Computing Environm...
summary
Describes a concrete stack: improved MPI Operator plus a Scheduling Framework plugin originally called KubeFlux now Fluence, integrating the Flux hierarchical scheduler into Kubernetes. Fixes include race-condition and DNS-flood patches that let declarative MPI bootstrap to 16,384 ranks on EKS. Benchmarks use AMG, LAMMPS and QMCpack on IBM Cloud/AWS and show Fluence-packed placements reduce variability and run some workloads up to three times faster than default kube-scheduler, with GPU and zone awareness.
AI-written summary · about
chapters (7)
- 0:00Introduction
- 0:21Traditional HPC workflow challenges
- 3:10Flux framework for workflows
- 5:16Scaling MPI with Kubeflow
- 6:42Fluence performance on Kubernetes
- 9:28Future work with Flux operator
- 10:12Q&A
related talks
28mKubeFlux: An HPC Scheduler Plugin for Kubernetes
06mProject Lightning Talk: Flux - What is it & What's New?
33mFlux Ecosystem Evolution
36mManaging Thousands of Clusters and Their Workloads with Flux
37mComputational Fluid Dynamics (CFD) Analysis With Kubernetes, Kubeflow, And OpenFOAM
19mKeynote: Building Bridges: Cloud Native and High Performance Computing
07mFluxCD - Gitops For All Sizes
42mIntro: Flux