FORD: Fast One-sided RDMA-based Distributed Transactions for Disaggregated Persistent...
summary
FORD uses one-sided RDMA to bypass remote CPUs and makes three concrete protocol changes: hitchhiker locking that batches read+lock into one RDMA to cut validation RTTs, coalescent commit that does in-place parallel undo-logging and updates primaries and backups in one round trip, and selective remote flash plus backup-enabled reads to reduce PM bandwidth pressure. Implementation on one client node and two PM nodes with Kiwi and TPC-C (8 warehouses, 112 coordinators) shows up to 2.3x throughput and up to 74.3% lower latency, and reduces the critical path from five RTTs to three.
AI-written summary · about
chapters (12)
- 0:00Intro
- 0:21Disaggregated Persistent Memory Memory disaggregation
- 1:57RDMA-based Distributed Transaction
- 2:35Transaction on Disaggregated PM
- 3:18State-of-the-art
- 4:22Challenge 1
- 8:05Overview
- 8:51Hitchhiked Locking Read and lock the read-write (RW) set in execution . Avoid subsequent locking and validations • No lock on the read-only data
- 9:49Coalescent Commit Commit all replicas together in one round trip - In-place update: Parallel undo logging in execution phase • Prevent reading partial updates: Control data visibility
- 11:50Selective Remote Flush Only issue remote flushes to backups after the final write • Ensure recoverability by backups • Reduce flushing round trips • Compatible with different flush primitives
- 14:00Evaluation Benchmarks
- 15:15Conclusion
related talks
19mFast Distributed Transactions for RDMA-based Disaggregated Memory
18mFARLock: Asymmetric RDMA Locking Made Fair
25mDisaggregating Persistent Memory and Controlling Them Remotely
14mODRP: On-Demand Remote Paging with Programmable RDMA
01h 09mAlgorithms for practical distributed agreement
28mPersistent Memory plus RDMA, new age remote device
23mMulti-Path Transport for RDMA in Datacenters
12mCharacterizing and Optimizing Remote Persistent Memory with RDMA and NVM