Algorithms for practical distributed agreement
summary
A practical RDMA-based state machine replication system called MU uses Infiniband one-sided operations to achieve a one-sided round trip commit in the common case (leader writes, followers are silent) and shows ~2.7x lower replication latency than prior RDMA systems. Permission revocation plus a pull-based heartbeat lets leader changes finish in under <1 ms failover (about 12× faster than the previous best). Implementation works on four-server tests with Herd, Redis, and LiquidBook and uses RDMA memory-region permissions to avoid extra CPU hops.
AI-written summary · about
chapters (23)
- 0:00Introduction
- 5:37Welcome
- 6:12What is agreement
- 12:00Practical applications
- 13:11Overhead
- 14:39Common Case Analysis
- 17:49Practical Algorithms
- 18:01The Plan
- 21:27RDMA basics
- 22:37Running time
- 23:27Background
- 29:27Diskpaxos
- 33:21Theoretical results
- 34:17Practical problems
- 36:52Multiinstance consensus
- 38:44Failure recovery
- 46:49Experimental setup
- 48:42Replication latency
- 49:51Latency comparison
- 51:20throughput
- 53:05Failover time
- 55:32Leaderless consensus
- 59:09Signatures
related talks
06mRDMA Programming Design and Case Studies - For Better Performance Distributed
16mFORD: Fast One-sided RDMA-based Distributed Transactions for Disaggregated Persistent...
39mTrust Deterministic Execution to Scale & Simplify Your Systems
18mRDMA is Turing complete, we just did not know it yet!
15mShiftLock: Mitigate One-sided RDMA Lock Contention via Handover
16mWaverunner: An Elegant Approach to Hardware Acceleration of State Machine Replication
26mBreaking Magical Barriers
19mFast Distributed Transactions for RDMA-based Disaggregated Memory