Near Optimal Latency Versus Cost Tradeoffs in Geo Distributed Storage
summary
Paper claims a concrete system-level win: by allowing erasure coding and reworking Paxos-style quorums the design called Pando cuts write latency from 120 ms to about 65 ms (within 10% of a one‑round protocol) while lowering storage cost. Evaluation on Azure compares to ePaxos and RS‑Paxos and shows Pando closes the gap to a theoretical lower bound far better (11x/13x gap improvements reported). Implementation details, quorum math, and cloud measurements are provided.
AI-written summary · about
chapters (12)
- 0:00Intro
- 0:21Distribute Web Servers for Interactive Latency
- 3:48How Do Existing Approaches Perform?
- 6:27Lowering Cost with Erasure Coding
- 7:34Recap of the Problem
- 8:24Paxos Review
- 9:17Quickly Executing 2 Phase Writes
- 11:43Write to All, Wait for Quorum
- 12:28Achieving 1 Site Intersection
- 13:16Evaluation: Proximity to Lower Bound
- 14:06Pando is Close to the Lower Bound
- 15:01Conclusion
related talks
12mEPaxos Revisited
22mLiberating Distributed Consensus
16mSwiftPaxos: Fast Geo-Replicated State Machines
20mHigh Throughput Replication with Integrated Membership Management
23mLightning Talks on Distributed Systems
24mStronger Semantics for Low-Latency Geo-Replicated Storage
28mDistributed Ledger
18mHigh Performance Memory For The World Computer