[Beowulf] IEEE 1588 (PTP) - a better cluster clock?
apittman at concurrent-thinking.com
Tue Jul 24 09:49:41 PDT 2007
On Tue, 2007-07-24 at 14:37 +0000, Patrick Ohly wrote:
> On Tue, 2007-07-24 at 15:15 +0100, Ashley Pittman wrote:
> [examples for the need of a more accurate clock]
> > But none of the ones you list are more than vaguely related to HPC.
> > The only thing I've found ntp to be lacking in is the ability to
> > compile
> > complex software over NFS on 2.6 kernels which have 64 bit timestamps,
> So you haven't used MPI tracing tools yet? This is very much HPC
> related. NTP is not good enough for that purpose, nor is a simple linear
> clock correction. One of the artifacts you get in the trace when the
> clocks of different processes are shifted against each other by more
> than the latency of short messages are "backwards" messages which are
> said to be received earlier than they were sent. Less obvious, but just
> as inaccurate are shifts in the other direction.
Perhaps I'm being somewhat naive in assuming that MPI_Wtime would take
care of this kind of problem within the job. I have used tracing tools
and I've never noticed the kind of artifact.
Don't get me wrong, ptp is clearly progress but it's not solving
something that has ever caused me significant problems and I wish people
would make more use of the tools that are already available.
More information about the Beowulf