DADI: Block-Level Image Service for Agile and Elastic Application Deployment
summary
A production block-based image service that replaces download+unpack with on‑demand reads, using a block-level image format plus an overlay block device and a seekable compressed blob format (zfile). Layer indexes merge into <4,500 entries (~72 KB) so reads hit millions/sec on one core, optional P2P balances traffic, and integrations include ext4/QEMU. Evaluation shows much lower cold-start latency and higher IO than overlayfs/LVM, achieving cold start 10,000 containers on 1,000 hosts in 4s.
AI-written summary · about
chapters (9)
- 0:00The container latency problem
- 2:10Limitations of current systems
- 3:46Block vs. file-based systems
- 6:22Architecture of DADI
- 8:53Efficient indexing and storage
- 11:32Seekable compression and layers
- 14:07P2P and shared storage scaling
- 16:25Performance evaluation
- 18:30Future directions
related talks
16mDADI @ Scale: Deploying Containers at Scale
25mBlazing-Fast Container Deployments: Image Caching With Block Storage
25mStreamlined Efficiency: Unshackling Kubernetes Image Volumes for Rapid AI Model...
18mPoby: SmartNIC-accelerated Image Provisioning for Coldstart in Clouds
22mMulti-Purpose Cloud Filesystem
23mOn-demand Container Loading in AWS Lambda
12mAdvancing Flash Storage
49mUntangle spaghetti of the bare metal images and modes