Secure Distributed Block Device for Edge and Cloud
summary
Mindcastle is a serverless distributed block store with compressed encrypted self-certifying chunks, HTTP-backed demand fetch, and incremental atomic updates built on an LSM-like data structure. A content-addressable layout uses a 260-bit AES key for values and root hashes to verify integrity. It drives tiny edge cameras and cloud spot-instance training: compression drops a worst-case 16 GB image to ~131 MB, and sequential access sustains ~1,000 images/sec across two GPUs on older hardware. Ongoing work adds rsync-style chunking and POSIX consistency.
AI-written summary · about
chapters (17)
- 0:00Intro
- 0:17What is Mindcastle?
- 0:47Edge Al smart cameras
- 1:27Containers on Edge
- 2:45ML Training in the Cloud
- 4:07VM-based Isolation
- 8:14Virtual disks are like databases!
- 9:40LSM-like "dubtree" data structure
- 15:00Evolving SWAP
- 16:30SWAP += encryption and distribution
- 17:17Structure of the Code
- 19:06Mindcastle.swap file example
- 21:28Merge avoidance
- 22:30Avalanching effect of merges
- 23:26Chunking scheme
- 25:02Current and future work
- 27:10Summary
related talks
18mOn Scalable Integrity Checking for Secure Cloud Disks
13mMlsDisk: Trusted Block Storage for TEEs Based on Layered Secure Logging
20mShared Security in the BLOCKCHAIN
01h 03mHow I Learned to Stop Worrying and Trust the Cloud
30mEncrypting Memory at Scale
30mScaling the Grail – Cloud-Native Computing on Encrypted Data using Carbyne Stack
35mConfidential Computing in the Control Plane (case study using Barbican)
44mAdvancements in Confidential Computing