Resource Orchestration of HPC on Kubernetes: Where We Are Now and the Journey Ahead!
Talkat Kubernetes Batch + HPC Day Europe 2022 · on Tuesday, 2022-05-17· at 13:55· Room 4F | Event Center·
summary
A NUMA-aware system is implemented: a NodeResourceTopology CRD plus a node agent (NFD topology updater or custom resource topology exporter) and a topology-aware scheduler plugin (filter + score) that reads per-NUMA capacity/available counters from the CRD. Components are released as images, a deployer CLI and an OpenShift operator; NFD gained updater support in v0.10 and topology code aligns with topology manager (introduced in Kubernetes 1.16, beta in 1.18). Results reduce topology-affinity errors but do not eliminate post-admission rejections due to kubelet reconciliation; a reserve plugin and tighter updater notifications are planned.
AI-written summary · about
chapters (11)
- 0:00Start
- 0:08Resource Orchestration of HPC on K8s Where We Are Now and the Journey Ahead!
- 1:59NUMA-unaware Scheduling - pain points
- 2:46Scheduler vs the kubelet
- 5:57Use cases 1. Workloads that require specialized hardware (HPC)
- 7:37NUMA-aware scheduling in Kubernetes
- 20:51Source code and container images
- 21:35Documentation and manifests
- 22:01The deployer toolkit
- 22:57Operator
- 23:44Future Plans
related talks
18mLet There Be Topology-Awareness in Kube-Scheduler!
26mResource as You Wish: Collaborative Reservation and Allocation b...
34mStreamline Your Cluster with Automated Node Labels,Taints, and Annotations
35mResource Management WG Deep Dive
47mNHD - A Topology-Aware Scheduler for K8s for Low-Latency & HPC Applications
34mKubernetes DRA Architecture: Scheduling, Status, and Topology at Scale
34mUnlocking Performance: Topology-Aware CPU Scheduling W...
09mMulti-Cluster Networking: A Technology-Agnostic Network Operator for K8 Federation