FARLock: Asymmetric RDMA Locking Made Fair
Talkat 20th USENIX Symposium on Operating Systems Design and Implementation (OSDI '26) · 2026-07-13..07-15·
summary
FARLock implements ticket and MCS-style handover plus separate local and remote queues to avoid RDMA loopback while enforcing strict arrival-order grants; it uses a Peterson's lock to atomically assign tickets to queue heads and a grouping optimization so one ticket serves consecutive same-queue requests. Evaluated on a 10-node ConnectX-3 RDMA cluster (9 remote clients, 12 threads each), it reports up to 62× lower local tail latency, up to 8× better than Alock, and speeds Sherman’s local tail by up to 12× and remote by up to 6×.
AI-written summary · about
chapters (13)
- 0:00FarLock and RDMA locking
- 1:47Local loopback in RDMA locks
- 2:50Fairness in asymmetric RDMA locks
- 4:14Combining queues with tickets
- 5:38ForkLock structure
- 6:25Acquiring and releasing locks
- 7:28FarLock in action
- 10:43Fairness grouping optimization
- 12:06Performance evaluation setup
- 13:07Tail latency results
- 14:11Throughput under contention
- 15:14FarLock with Sherman B+ tree
- 16:40Summary and code
related talks
15mShiftLock: Mitigate One-sided RDMA Lock Contention via Handover
16mFORD: Fast One-sided RDMA-based Distributed Transactions for Disaggregated Persistent...
15mCitron: Distributed Range Lock Management with One-sided RDMA
19mFast Distributed Transactions for RDMA-based Disaggregated Memory
15mSecuring RDMA for High-Performance Datacenter Storage Systems
19mFineMem: Breaking the Allocation Overhead vs. Memory Waste Dilemma in Fine-Grained...
14mFast and Scalable In-network Lock Management Using Lock Fission
15mFlor: An Open High Performance RDMA Framework Over Heterogeneous RNICs