DADI @ Scale: Deploying Containers at Scale
summary
DADI is a block-level image system that replaces full image download and unpack with on-demand block reads over a 3D P2P topology, using an overlay block device format plus a random-access online compression (refill) and trace-based prefetching. Deployed across Alibaba to support Kata, Firecracker and gVisor, it achieves ~0.7s cold-start latency at scale versus 1.5–2.3s for alternatives, runs 10,000 containers on 1,000 hosts in experiments, and ships open source with contributions to containerd and Harbor.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 1:09Why DADI is needed
- 3:17Existing solutions limits
- 4:42File system vs block images
- 7:08Container image structure
- 8:13DADI layered format
- 10:42P2P transport and caching
- 11:47I/O path and startup
- 12:32Performance evaluation
- 13:35Scalability test 10k containers
- 14:21Comparison with overlayfs
- 14:42Open source future
related talks
21mDADI: Block-Level Image Service for Agile and Elastic Application Deployment
25mBlazing-Fast Container Deployments: Image Caching With Block Storage
36mcontainerd Deep Dive
18mPoby: SmartNIC-accelerated Image Provisioning for Coldstart in Clouds
36mExtending containerd
01h 03mContainers for dotnet developers
34mCDI: The Future of Specialized Hardware in Containers
34mContainerd: Project Update and Deep Dive