Slinky: Slurm in Kubernetes, Performant AI and HPC Workload Management in Kubernetes
Talkat KubeCon + CloudNativeCon Europe 2025 · on Wednesday, 2025-04-02· at 12:00· Level 1 | Hall Entrance S10 | Room A·
summary
Slinky is a toolkit (open source Apache 2 on GitHub) that embeds Slurm into Kubernetes via a Slurm operator, a slurm bridge, a client library and a Prometheus exporter. Slurm claims scale beyond 15,000 nodes and can launch 10,000+ node simulations in under 10s; Slinky v1 appeared last November, v2 just released, v3 planned for June. The operator runs slurmd pods and autoscaling; the bridge turns multi-node pod groups into Slurm placeholder jobs and schedules them, currently assuming exclusive node ownership and limited safe mixing with native pods.
AI-written summary · about
chapters (7)
- 0:00Slinky and SLURM intro
- 3:56Slinky components and releases
- 12:52Converging HPC and cloud native
- 14:58Slinky operator architecture
- 20:18Slinky bridge design
- 26:44Slinky bridge demo
- 29:56Future work and questions
related talks
05mHow To Run Kubernetes Pods On My Slurm-Based HPC Center
12mLightning Talk: {”spec”: “nodeSelector”: "EuroHPC"}
32mPods Everywhere! InterLink: A Virtual Kubelet Abstraction Streamlining HPC Resource Exploitation
37mUsing SLOs for Continuous Performance Optimizations of Your K8s Workloads
01h 05mIntro + DeepDive: SIG Scalability
25mAn SLO-Driven Approach to Enhance Kubernetes Cluster Reliability
34mDeep Dive: Scalability SIG
28mKubeFlux: An HPC Scheduler Plugin for Kubernetes