Benchmarking HPC workloads on OpenShift
summary
OpenShift runs as a proof‑of‑concept HPC platform on a 32‑node cluster compared with identical bare‑metal nodes, using MPI Operator, Multus CNI and OpenShift Container Storage plus a Kubernetes Go client and a “matrix benchmarking” Python tool that collects and visualizes >180 measurements for Specfem and many for Gromacs. Results show SDN adds noticeable latency; bypassing it with Multus or host networking yields near bare‑metal performance, Gromacs needs SDN bypass for multi‑node scaling, and Specfem shows <6% overhead up to 16 nodes and <4% for larger problem sets.
AI-written summary · about
chapters (12)
- 0:00Start
- 0:46Why We Want To Run Hpc Applications on Openshift
- 3:25Proof of Concept Platform
- 4:25Openshift Container Storage
- 5:47Mpi Operator
- 7:37Run a Specfam on Openshift
- 9:34Prepare a Custom Resource Definition
- 10:11Parallel Execution
- 12:01Matrix Benchmarking
- 14:38Benchmarking Results
- 15:38Mpi Collective Operations
- 16:07Spectfem Results
related talks
30mOptimizing costs across several OpenShift clusters
44mPerformance and Scalability Tuning Kubernetes for OpenShift and Docker
31mPod scaling ad infinitum!!
13mOptimizing Workload Management in a Multi-Cluster Environment: A Case Study
37mComputational Fluid Dynamics (CFD) Analysis With Kubernetes, Kubeflow, And OpenFOAM
10mOvercome Interconnect Bottlenecks in HPC Clusters
46mBuilding Supercomputers Using Kubernetes and OpenHPC
47mCloud Native and Sustainable, Reproducible Scientific Computing