Sustainability by Design: Leveraging DRA for Energy-Efficient…
summary
Demo shows a proof‑of‑concept where vendors expose TDP and runtime watts via a DRA driver so the scheduler can pick nodes labeled performance/balanced/efficient across a four‑node cluster, Kepler collects GPU joules for all vendors, and KEDA enforces an energy budget by scaling pods down when average GPU joules hits 150. Kubernetes 1.26 marks DRA stable, the proposal adds TDP fields to resource slices, and the live demo successfully schedules to energy‑efficient device classes and triggers autoscale actions.
AI-written summary · about
chapters (14)
- 0:00Sustainability in Kubernetes
- 1:03Energy efficiency priority
- 3:11Hardware and workload factors
- 4:37Dynamic Resource Allocation
- 5:41DRA for scheduling
- 7:03DRA for provisioning
- 7:47Kepler for energy metrics
- 9:12Vendor DRA energy data
- 10:37Demo of DRA energy enhancements
- 15:32Kepler’s DRA energy metrics
- 16:15KEDA energy scaling
- 16:56Cloud-native sustainability loop
- 17:58Contributing to sustainability
- 19:02Q&A on energy prediction
related talks
31mMore Nodes, More Problems: Solving Multi-Host GPU/TPU Scheduling With Dynamic Resource Allocation
20mEnergy Efficient Placement of Edge Workloads
20mMonitoring Energy Consumption and Building Energy Efficient Systems the Cloud Native Way
20mKepler - Optimized power usage monitoring for Kubernetes
59mHow Kubernetes optimisation can combat climate change
35mSustainability Research the Cloud Native Way
47mBuilding sustainable architectures with Kubernetes
28mExploring Kepler's Next Chapter: Achieving Cloud Native Sustainability With MCP Integration