[mpich-discuss] Clock synchronization

Anthony Chan chan at mcs.anl.gov
Sat Apr 2 22:46:31 CDT 2011


MPE logging synchronizes PMPI_Wtime since MPI implementation
does not, so you could take a look at the code at
src/mpe2/src/logging/src/clog_time.c and clog_sync.c.
The code assumes there is no drift in MPI_Wtime, it only
calculates an offset....  There are several algorithms,
the accurate one isn't scalable, the scalable one is less accurate.
The accurary of the offset generally deteriorates with the increasing
number of nodes, IIRC.

A.Chan

----- Original Message -----
> [please keep mpich-discuss cc'ed].
> 
> There is no real communication during synchronization. Zero-byte
> messages are exchanged. The algorithm used is logarithmic with respect
> to the number of processes.
> 
> -- Pavan
> 
> On 04/02/2011 02:52 AM, Ravi Prakash wrote:
> > Thanks Pavan! :) Do you know if synchronization is done by splitting
> > the
> > communication time in half or somesuch algorithm?
> >
> > Cheers
> > Ravi
> >
> > ------------------------------------------------------------------------
> > *From:* Pavan Balaji <balaji at mcs.anl.gov>
> > *To:* Ravi Prakash <ravihoo at ymail.com>; mpich-discuss at mcs.anl.gov
> > *Sent:* Friday, April 1, 2011 11:33 PM
> > *Subject:* Re: [mpich-discuss] Clock synchronization
> >
> >
> > Assuming TCP/IP communication and a 10-20 machines, once you do an
> > MPI_Barrier, I don't expect the time difference between the machines
> > to
> > be more than a few 10s of microseconds, if not lesser.
> >
> > -- Pavan
> >
> > On 04/01/2011 11:28 PM, Ravi Prakash wrote:
> >  > Hi,
> >  >
> >  > I'm a n00b and I'm wondering if there is any documentation on
> >  > what clock
> >  > synchronization algorithm MPICH2 uses. I need to trigger events
> >  > at the
> >  > same time across multiple machines. What accuracy am I looking
> >  > at?
> >  >
> >  > My machines are Windows Vista boxes with Dual dual-core Xeons
> >  > connected
> >  > by an ethernet 100Mbps LAN. Please let me know if there are any
> >  > more
> >  > relevant factors needed for an approximate figure. Is it a few ms
> >  > / 10's
> >  > of ms / 100's of ms / god forbid seconds? :)
> >  >
> >  > Cheers
> >  > Ravi
> >  >
> >  >
> >  >
> >  > _______________________________________________
> >  > mpich-discuss mailing list
> >  > mpich-discuss at mcs.anl.gov <mailto:mpich-discuss at mcs.anl.gov>
> >  > https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss
> >
> > --
> > Pavan Balaji
> > http://www.mcs.anl.gov/~balaji
> >
> >
> 
> --
> Pavan Balaji
> http://www.mcs.anl.gov/~balaji
> _______________________________________________
> mpich-discuss mailing list
> mpich-discuss at mcs.anl.gov
> https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss


More information about the mpich-discuss mailing list