This preview shows page 1. Sign up to view the full content.
Unformatted text preview: hat you can just assign every event in the distributed system a real-‐time stamp. But that doesn’t work, since clocks aren’t perfect – they need to be synchronized, they drift with respect to each other, and the only thing we have to do the re-‐synchs, are unreliable messages of uncertain performance. Lets assume for simplicity that there are some nodes in the system have well-‐
defined wall-‐clock time, that is, they have a GPS unit. Even with a GPS, you need to distribute the GPS information over the network to the other nodes in the system. How close can we bound the clocks on those other systems? Option 1: if there’s an upper bound on the message time, can synch to 2max RTT but there (often) isn’t an upper bound in practice (or its absurdly high). Even on a LAN, RTT > 1M instructions Option 2: keep fetching the time, estimate relative rate of skew of the local clock. Can give a somewhat tighter bound. Do either of these solve the problem with the makefile timestamps? <We’ll se...
View Full Document
This document was uploaded on 04/04/2014.
- Spring '14