uKharon: A Membership Service for Microsecond Applications
summary
uKharon is a membership service that uses RDMA to push membership and failover to microsecond scales: a multi-level failure detector (kernel crash notifications + an RDMA-exposed heartbeat counter), a Paxos-like consensus implemented with one-sided RDMA CAS, and tiny leases clients renew in the background (about 20µs leases). Built uKharon-KV removes replicated GETs vs HERD, keeps PUT latency comparable, and drives end-to-end failover to 50–53µs, a roughly 10× improvement over RDMA SMR approaches.
AI-written summary · about
chapters (12)
- 0:00Intro
- 0:42The data center is a zoo!
- 1:17How do we usually deal with failures?
- 4:17Remote Direct Memory Access (RDMA)
- 5:07Microsecond-scale failure detectors
- 8:29Microsecond-scale replication
- 9:32One-sided Paxos
- 10:33Membership invalidation
- 12:02Microsecond-scale leases
- 14:05Evaluation: setup
- 14:31Evaluation: Replicating a KV-Store
- 17:43Conclusion
related talks
01h 09mAlgorithms for practical distributed agreement
14mUnderstanding RDMA Microarchitecture Resources for Performance Isolation
13mLive in the Express Lane
15mHydra : Resilient and Highly Available Remote Memory
15mReplicating Persistent Memory Key-Value Stores with Efficient RDMA Abstraction
57mReliability Nirvana
18mFARLock: Asymmetric RDMA Locking Made Fair
24mPLOVER: Fast, Multi-core Scalable Virtual Machine Fault-tolerance