Serving HTC Users in Kubernetes by Leveraging HTCondor
Talkat KubeCon + CloudNativeCon North America 2019 · on Thursday, 2019-11-21· at 15:20· Room 6E - San Diego Convention Center, Upper Level·
summary
Describes deploying HTCondor inside Kubernetes to serve true HTC workloads—persistent central manager plus scalable execute nodes—claiming production use across Open Science Grid at ~100,000 cores and a one-off pool of 270,000 CPUs / 350 PFLOP. Uses Singularity for user containers (including Singularity-in-Docker nested containers when kernels allow), injects Condor via an init container into user images, supports preemptable jobs that reschedule transparently, and offers three architectures (Condor services outside, inside, or mixed with Kubernetes) that materially raise cluster utilization.
AI-written summary · about
chapters (15)
- 0:00Intro
- 1:32Ingenious Parallelism
- 2:27Example HTC problems
- 3:09Example HTC resource
- 3:36Example HTC users
- 4:15HTC and Kubernetes
- 7:09HTCondor Architecture
- 10:00Using HTCondor with Kubernetes
- 12:18HTC Users and Containers
- 12:57HTCondor and Containers
- 14:54Explicit provisioning
- 17:17Opportunistic use
- 19:33To conclude
- 20:18Acknowledgments
- 26:57Nested containerization
related talks
19mKeynote: Building Bridges: Cloud Native and High Performance Computing
37mRunning Mixed Workloads on Kubernetes at the Institute for Health Metrics and Evaluation
41mBuilding & Operating a Large-scale HPC AI Cluster on Kubernetes
30mTuning and automating for Telco 5G containerized workloads
20mA full-scenario colocation of workloads based on Kubernetes
08mLightning Talk: Orchestrating Kubernetes Clusters on HPC Infrastructure
47mCloud Native and Sustainable, Reproducible Scientific Computing
46mBuilding Supercomputers Using Kubernetes and OpenHPC