Hamband: RDMA Replicated Data Types
summary
First hybrid replicated data types for RDMA that implement synchronization groups (strongly connected components of a conflict graph) with a single-leader synchronization log, per-node dependency maps, and a summarization groups mechanism that collapses reducible updates into a single memory location instead of per-node logs. Uses single-writer sharded RDMA memory, a reliable RDMA broadcast for crash handling, and fine-grained synchronization to increase parallelism. Evaluation shows much lower latency and higher throughput and scalability than a sequentially consistent message-passing baseline; follower crashes barely reduce throughput while leader crashes hurt more.
AI-written summary · about
chapters (9)
- 0:00Start
- 2:12Hybrid Consistency Models
- 3:29Example of Conflict
- 4:29Dependency
- 5:00Well-Coordinated Execution
- 9:09Track Dependencies
- 12:23Experimental Results
- 13:28Fault Tolerance Result
- 14:28Conclusion
related talks
01h 09mAlgorithms for practical distributed agreement
19mFast Distributed Transactions for RDMA-based Disaggregated Memory
16mFORD: Fast One-sided RDMA-based Distributed Transactions for Disaggregated Persistent...
18mRDMA is Turing complete, we just did not know it yet!
15mReplicating Persistent Memory Key-Value Stores with Efficient RDMA Abstraction
17mDuhu: Shared Disaggregated Memory for Distributed Data Processing Frameworks
15mOne-sided RDMA-Conscious Extendible Hashing for Disaggregated Memory
28mPersistent Memory plus RDMA, new age remote device