The Coral Reef Cloud: Autoscaling, Reservation, and Preemption
summary
OpenStack Blazar reservations get a temporal lease so platforms can guarantee future capacity while the cloud fills gaps with preemptible instances that Blazar can terminate just before a reservation starts. Kubernetes (Magnum) and Slurm integration use the cluster autoscaler pattern to spin VMs up and down from the cloud image catalog; Slurm boots configured images as nodes and releases them when idle. A prototype for the IRIS/Cambridge deployment demonstrates higher utilization and a proposed flexible reservation API plus the cluster autoscaler workflow under development upstream.
AI-written summary · about
chapters (10)
- 0:00Introduction to Coral Reef Cloud
- 2:10HPC 2.0 and converged cloud
- 6:01Quotas and fair share challenges
- 9:13Blazar reservations for guarantees
- 11:04Auto scaling platforms explained
- 12:28Slurm integration with cloud
- 14:55Preemptible instances concept
- 16:37Building the Coral Reef model
- 18:05Next steps for Blazar
- 19:32How to get involved
related talks
47mGet off my GPU! Better Sharing of AI Accelerators in OpenStack with Blazar Reservations
06mAutoscaling GPU Clusters Anywhere — Hyperscalers, Neoclouds & Ba...
05mSponsored Keynote: The Cloud (And Your $$$) is Not Infinite: Dynamic Scheduling at Every Layer
39mCluster API & Cluster Autoscaler
31mAccelerate AI adoption: automated cloud efficiency
21mWhy should you bother about cluster autoscaling- Karpenter
36mCluster Autoscaler Evolution
31mDeclarative Underlays: Scaling Purpose-Built Infrastructure(Clusters)...