One-sided RDMA-Conscious Extendible Hashing for Disaggregated Memory
summary
RACE hashing is an RDMA-conscious extendible hash table that executes every index operation using one-sided RDMA verbs, avoids remote compute by keeping all state in the memory pool, and provides a lock-free remote concurrency control to let multiple clients operate the same table without remote locks. It implements extendible remote resizing that avoids an extra RDMA to fetch directories, caches directory entries on clients with lightweight validation, and in YCSB hybrid workloads outperforms prior distributed in-memory hash indexes by 1.4–13.7×.
AI-written summary · about
chapters (16)
- 0:00Intro
- 0:19Disaggregated Memory
- 0:54Hashing Indexes in Local Memory
- 1:30Hashing Indexes in Disaggregated Men
- 1:55Hash Collision
- 4:04Challenge 3: Extendible Resizing
- 5:25Challenge Summary
- 6:12Architectural Overview
- 6:42One-sided RDMA-Conscious (RAC) Subtabl
- 8:11Lock-free Remote Concurrency Control
- 10:19Extendible Remote Resizing
- 12:06Experimental Setup
- 12:35Insertion
- 13:25Update
- 13:35YCSB Hybrid Workloads
- 13:47The Stale-read Client Directory (SRCD)
related talks
14mSMART: A High-Performance Adaptive Radix Tree for Disaggregated Memory
24mWrite-Optimized Dynamic Hashing for Persistent Memory
13mDMTree: Towards Efficient Tree Indexing on Disaggregated Memory via Compute-side...
18mCuckoo for Clients: Disaggregated Cuckoo Hashing
19mFast Distributed Transactions for RDMA-based Disaggregated Memory
17mDuhu: Shared Disaggregated Memory for Distributed Data Processing Frameworks
15mROLEX: A Scalable RDMA-oriented Learned Key-Value Store for Disaggregated Memory Systems
14mSEPH: Scalable, Efficient, and Predictable Hashing on Persistent Memory