Dynamic GPU Provisioning in Kubernetes With Composable Disaggregated...
Talkat Open Source Summit Japan + AI_dev: Open Source GenAI & ML Summit + Automotive Linux Summit 2025 · 2025-12-08..12-10·
summary
CoHDI adds dynamic GPU provisioning to Kubernetes by exposing a resource-pool device list and pausing scheduling until GPUs are attached, using three components: Composable DRA Driver (publishes pooled GPUs), Dynamic Device Scaler (increments/decrements requested GPU counts) and Composable Resource Operator (calls the CDI manager to attach/detach). Demo shows GPUs attach from the pool and pods move from Pending to Running, attachment latency around 10–20 seconds, and scheduler device binding conditions prevents container failures.
AI-written summary · about
chapters (23)
- 0:00Dynamic GPU provisioning intro
- 1:11GPU resource challenges
- 2:14Composable disaggregated concept
- 3:40CDI software stack
- 4:25Dynamic GPU scaling
- 5:31Kubernetes GPU limits
- 7:00Kubernetes Dynamic Resource Allocation
- 9:34GPU resource pool problems
- 11:41Kodi open source project
- 12:47Kodi integration design
- 14:57Dynamic provisioning architecture
- 16:02GPU provisioning flow
- 18:36Composable DA driver details
- 20:25Composable resource operator
- 22:55Dynamic device scaler
- 25:03Scaling challenges
- 26:06Scheduler enhancements for GPU
- 28:16Demo setup overview
- 30:54Demo walkthrough
- 36:28Summary and future challenges
- 37:10Q&A
- 37:36Performance and scheduling
- 40:59Closing remarks
related talks
38mDynamic Scaling of GPUs for Container Apps with Composable Disaggregated...
30mCoHDI: A CNCF Sandbox Project for Dynamic Hardware Composability in Kubernetes
04mFrom Idle to Ideal: Cross‑Cluster GPU Sharing with CoHDI
05mProject Lightning Talk: CoHDI: Dynamic Hardware Composition for AI‑Era Kubernetes…
30mRunning AI Workloads in Containers and Kubernetes
34mDeclarative Device Virtualization: Orchestrating GPUs & Hardware in Cloud Native Environments
05mTowards a Sustainable World Realized by Kubernetes and Composable...
24mFit-to-Serve: How a New DRA Capability for Dynamic Device...