Containerization of Parallel MPI and Accelerated Applications on HPC Systems
summary
Presents measurements showing Singularity containers run MPI and GPU HPC jobs with negligible overhead compared to bare metal. Benchmarks include Quantum Espresso on 128 nodes (6,144 cores) using Intel and OpenMPI builds, and TensorFlow GPU jobs on NVIDIA K80s accessed via Singularity’s --nv binding using vendor-optimized Docker images converted to Singularity and standard batch schedulers.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 1:27HPC cluster
- 5:26Containers in HPC
- 6:52HPC container tech
- 9:18Singularity features
- 12:52Singularity workflow
- 16:04MPI parallel apps
- 18:51GPU accelerator support
- 19:12MPI app performance
- 25:40MPI test results
- 26:01GPU app performance
- 28:54Next steps
related talks
08mContainerization – It’s Not Just for CPUs Anymore: Introducin...
57mMoby Dock in the Wind Tunnel
18mImpossible is Nothing! The 10x hpothesis
42mBuilding a green supercomputer in the cloud
38mDynamic Scaling of GPUs for Container Apps with Composable Disaggregated...
32mA Deep Dive on Supporting Multi-Instance GPUs in Containers and Kubernetes
39mIntro + Deep Dive: SIG Scalability
23mOffload Annotations: Bringing Heterogeneous Computing to Existing Libraries...