FARLock: Asymmetric RDMA Locking Made Fair

Distributed locking is essential for coordinating access to shared resources in modern RDMA-based distributed systems. While state-of-the-art RDMA locks can deliver highperformance by introducing asymmetry (i.e., treating requests that are local and remote to the lock differently), they often trade off fairness as they fail to grant locks in the expected first-come first-serve manner. This can lead to long delays for critical tasks, missing service level objectives.

We present FARLock, a fast and fair RDMA lock to solve this problem. Drawing inspiration from ticket and MCS locks, FARLock employs tickets and MCS-style handover to ensure that locks are granted strictly by arrival order. Through careful coordination between request queues and the ticket, FARLock provides strong fairness semantics with high performance. Our evaluation on a range of workloads shows that FARLock guarantees fairness and achieves lower latencies compared to prior state-of-the-art. Incorporating FARLock in a recent RDMA-based distributed indexing solution improves its query tail latencies.

Recent work [4] mitigates this issue by introducing asymmetric RDMA locking where lock requests targeting a lock that is local to the requesting thread are handled separately from those that target a lock stored on another server. This way, local requests can perform local CPU-based atomics without involving expensive RDMA operations, which are only required by remote requests. Given the large latency gap between local and RDMA memory accesses (∼100ns vs. µs-level), this can greatly improve overall lock operation performance. As we elaborate in Section 2.3, however, it sac-