Fast Distributed Transactions for RDMA-based Disaggregated Memory
summary
HDTX is a distributed transaction system for RDMA disaggregated memory that claims up to 88.3% and 72.1% lower latency and up to 2.08× and 84.7% higher throughput versus FaRM and FORD by combining three concrete techniques: a fast commit protocol that merges execution/locking/validation/commit into two network round trips, an RDMA-enabled offloading mechanism that uses NIC work-queues and RDMA write/atomic primitives to copy redo logs on memory nodes without CPU, and a decentralized priority-based locking queue that shortens tail latency; evaluated on a five-server cluster with two memory nodes across TPC-SA, SmallBank and TATP workloads, the system shows large latency and throughput gains, especially under read–write conflict workloads.
AI-written summary · about
chapters (11)
- 0:00HDTX for RDMA disaggregated memory
- 1:22Traditional RDMA transaction challenges
- 3:08State-of-the-art DM transaction processing
- 5:11HDTX system overview
- 6:17Fast commit with redo logs
- 8:23Validation and commit efficiency
- 9:26ARIC data synchronization offload
- 11:54Priority-based locking for latency
- 14:27Experimental setup and benchmarks
- 16:55HDTX performance evaluation
- 17:39Conclusion and future work
related talks
16mFORD: Fast One-sided RDMA-based Distributed Transactions for Disaggregated Persistent...
18mRDMA is Turing complete, we just did not know it yet!
23mTeRM: Extending RDMA-Attached Memory with SSD
17mDuhu: Shared Disaggregated Memory for Distributed Data Processing Frameworks
12mDirect Access, High-Performance Memory Disaggregation with DirectCXL
12mCharacterizing and Optimizing Remote Persistent Memory with RDMA and NVM
25mDisaggregating Persistent Memory and Controlling Them Remotely
23mMulti-Path Transport for RDMA in Datacenters