Aurogon: Taming Aborts in All Phases for Distributed In-Memory Transactions
summary
Distributed in-memory system Aurogon implements 2LClock (two-layer RDMA-aware clock sync), adaptive request deferral for hotspot reads, and pre-attached writes that install a dummy version to eliminate a vulnerable interval and save one network round. Evaluated on a five-node RDMA cluster with TPC-C and YCSB/SSB, Aurogon raises throughput up to 4.1x, cuts aborts by up to 73%, cuts latency up to 86%, and reduces clock sync error (avg 51%, p99 2.3x) versus prior systems.
AI-written summary · about
chapters (16)
- 0:00Intro
- 0:24Transactions on rise
- 1:25Timestamp ordering protocol
- 2:35Problem of T/O protocols
- 3:15Existing solutions
- 3:56Inaccurate distributed clocks
- 4:51Nonuniform data access latency
- 5:31Requests with data dependency
- 6:14Challenges
- 7:52Two-layer clock synchronization
- 9:58Adaptive request deferral
- 10:53Pre-attaching dependent requests
- 11:51Evaluation Overview
- 12:16Evaluation: TPC-C benchmark
- 12:52Evaluation: YCSB benchmark
- 13:34Other evaluations
related talks
19mFast Distributed Transactions for RDMA-based Disaggregated Memory
15mCitron: Distributed Range Lock Management with One-sided RDMA
18mFARLock: Asymmetric RDMA Locking Made Fair
01h 09mAlgorithms for practical distributed agreement
15mADOC: Automatically Harmonizing Dataflow Between Components in Log-Structured Key-Value
16mScaleDB: A Scalable, Asynchronous In-Memory Database
17mMako: Speculative Distributed Transactions with Geo-Replication
45mBeyond Machine Boundaries: Experience with Memory Disaggregation