Propagating Programming Paradigms: Lifting High Performance Compute Into K8s
summary
Argues Kubernetes keeps useful pieces like device plugins and pluggable infrastructure but lacks HPC guarantees: requests node and network behavior consistency for a workload, software/firmware version affinity, and better checkpointing. Describes a new CPU manager control plane that pins cores and supports mixed/isolation pools, showing over threefold speedup in tests and raising CPU utilization from ~48% to >70%. Calls out the kubelet’s herd of managers as the main architectural pain point.
AI-written summary · about
chapters (17)
- 0:00Purpose of this talk
- 1:46HPC system assumptions
- 3:58Cloud vs HPC differences
- 6:06K8s device plugins
- 7:52Cloud paradigms to preserve
- 9:18Missing guarantees in Kubernetes scheduling
- 11:03CPU management for speed
- 12:50Need for compute math
- 14:13Correctness vs efficiency
- 14:41Q&A: CPU management
- 15:25Q&A: Singularity runtimes
- 16:29Q&A: Performance trade-offs
- 17:36Q&A: Resource awareness
- 18:24Q&A: HPC storage in K8s
- 19:51Q&A: Accelerators and drivers
- 21:39Q&A: Network behavior
- 22:46Q&A: Checkpointing and data
related talks
34mUnlocking Performance: Topology-Aware CPU Scheduling W...
23mThis is The Way: A Crash Course on the Intricacies of Managing CPUs...
46mBuilding Supercomputers Using Kubernetes and OpenHPC
07mKeynote: High Performance Computing on Google Kubernetes Engine
19mKeynote: Building Bridges: Cloud Native and High Performance Computing
38mKubernetes Networking Infrastructure Offload
24mUsing Kubernetes with Data Processing Units to Offload Infrastructure
29mKubernetes Workload Resiliency in Action: Beyond Basics