This preview shows page 1. Sign up to view the full content.
Unformatted text preview: e later on with Google’s Spanner a system that takes the approach -‐-‐ events are only located within a window of time of several milliseconds, and the system needs to be able to handle that.> If no external events, don’t need physical time, so Lamport argued for “logical clocks”. Very simple version of logical clocks: complete centralization: send every message to a central arbiter, which picks an order for all messages, and then system can deliver them in that order. Problems with centralization? Obviously scale – it wouldn’t work for the FB example we started with. But also, need to somehow recover if/when the central arbiter fails. Is a central arbiter required for consistent delivery everywhere? Lamport proposes an alternate algorithm. Introduce “space time diagram”: 3 processes, each does an RPC p1, p2, p3: p2-‐>p1, p1-‐>p3, p3-‐>p1, p3-‐>p2, p1-‐>p2, p1-‐>p3 p2 does RPC to p1; p1 does RPC to p3; p3 replies to p1; p3 does RPC to p2; p1 replies to p2; p1 does RPC to p3 Example #5:...
View Full Document
This document was uploaded on 04/04/2014.
- Spring '14