Easy Geo-Redundancy with MARS
summary
MARS is a production kernel module that provides long‑distance geo‑redundancy for ~15,000 servers, millions of customers and multiple petabytes of data, running background migration and live logical‑volume migration. It guarantees durability by appending writes to a persistent transaction log, reorders writes for locality, and ships them via up to 8 parallel TCP connections (separating metadata, replication and sync traffic) to avoid DRBD’s single‑stream bottleneck. It’s proven in service, uses CRC32/MD5 checksumming and is being prepared for upstreaming.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 0:54Agenda
- 1:35Downstream
- 4:43SLA
- 10:43Lab Testing
- 12:24Transaction Log
- 13:54Internals
- 17:33Current Status
- 22:26Architecture Differences
- 23:34Device Member
- 25:27Bricks
- 29:55Siblings
- 33:12Future plans
- 41:11Conclusion
related talks
31mFarm Convergence with DRBD
15mMigrOS: Transparent Live-Migration Support for Containerised RDMA Applications
25mEvolution of Geo-replication in Gluster
18mClient Driven Geo-Redundancy for Ceph Object Storage
24mMD cluster raid
24mPLOVER: Fast, Multi-core Scalable Virtual Machine Fault-tolerance
18mLeveraging Multiple Regions to Improve Site Reliability:
17mMAPX: Controlled Data Migration in the Expansion of Decentralized Object Based Storage