Explore a groundbreaking approach to fault-tolerant replication for microsecond-scale applications in this 20-minute conference talk from OSDI '20. Dive into the Mu system, which achieves replication in less than 1.3 microseconds and fail-over in under a millisecond, significantly outperforming previous systems. Learn how Mu leverages RDMA to implement state machine replication with strong consistency, addressing challenges such as concurrent leaders and log garbage collection. Discover the system's performance benefits when applied to various applications, including a financial exchange app, Redis, Memcached, and HERD. Gain insights into the innovative combination of RDMA permissions and distributed algorithmic design that enables Mu's exceptional speed and efficiency in replicating microsecond-scale applications.
Overview
Syllabus
Intro
Does Consensus Have to Be Slow?
Why Care About Microsecond Consensus?
Outline
Background: RDMA
Background: State Machine Replication
Mu Roles
Mu Common Case Replication
Common Case Replication: Intuition
Mu Background Plane
Mu Failure Detection
Evaluation: Setup
Evaluation: Replication Latency
Evaluation: Fail-over time
Conclusion
Taught by
USENIX
