[Planetlab-users] No guarantees regarding monotonically increasing time?

[Rob Sherwood]

On Fri, Aug 31, 2007 at 03:05:09AM -0500, Patel, Jay A wrote:
> Hi,
>  
> A few hours ago -- one of the nodes I was running an experiment on
> (system18.ncl-ext.net) experienced a negative movement in time. I
> assumed that PL nodes adjusted their clocks using NTP. Doesn't NTP
> guarantee monotonically increasing time? Or, did I just observe a manual
> time adjustment?
>  
> Has anyone experienced this?

Yes, this is unfortunately a known fact about planetlab.  It's documented in:


	A Platform for Unobtrusive Measurements on PlanetLab,
	Rob Sherwood and Neil Spring
	USENIX Workshop on Real, Large Distributed Sytems, November 2006.
	http://www.cs.umd.edu/~capveg/worlds06.paper.pdf

and

	Using PlanetLab for Network Research: Myths, Realities, and Best Practices,
	Neil Spring, Larry Peterson, Andy Bavier, and Vivek Pai
	ACM SIGOPS Operating Systems Review, 40:1, pages 17--24, January 2006.
	http://www.cs.umd.edu/~nspring/papers/osr.pdf

among other places.  The best I can recommend is that you make sure
that your program doesn't logically depend on monotonic time
increases.  For accurate timing measurements, I am exploring use of the
rdtsc assembly instruction that returns the number of CPU ticks.  It is
guaranteed to increases monotonically under certain conditions, and is
relatively straight forward to convert to wallclock time if you know
the machine's CPU speed.

Hope this helps,

- Rob
.