[Planetlab-users] Synchronizing clock and
measuring one-way latency
vivek at CS.Princeton.EDU
Tue Jul 5 18:01:14 EDT 2005
Neil Spring wrote:
> I figured Jawwad would get to that -- if he has to synchronize in the
> wide area, he'll likely want a somewhat stable source of time
> increments, like the tsc, rather than the NTP-adjusted kernel clock.
OK, got it. I just wanted to make it clear that Aki's code wasn't ever
intended for anything more than local time difference measurements.
> Are you starting to think that, in the post-v2 era, NTP is doing a
> good-enough job keeping clocks in sync? It seems plausible.
The widespread, systematic clock problems are gone, for the most part.
Those seemed mostly due to PIO-based disk access causing clock
interrupts to get dropped.
Looking at the Drift values in CoMon, there's about 6 nodes with more
than a few seconds of drift. My guess is that NTP is filtered at those
sites, so someone will have to beat on PIs to provide better local NTP
info. There was a wave of that a few months ago, and almost all nodes
had reasonably synchronized clocks for a while.
NTP obviously isn't good enough for tight coupling between different
nodes, but for applications that don't need time accurate to more than
a few seconds, the defaults are pretty good these days.
More information about the Users