AMBER Archive (2009)

Subject: Re: [AMBER] Number of Cycles

From: Robert Duke (rduke_at_email.unc.edu)
Date: Tue Dec 15 2009 - 11:29:22 CST

You pretty much have the major points, and yes, writing restarts does whack
performance at high processor count. Depending on the system you are
running on, the impact can get noticeable. Writing binary mdcrds can also
whack performance on some systems - the earlier cray xt machines in
particular, because you could not use the iobuf libraries. There are lots
of little gotcha's in all this stuff, and I really can't control everything;
a lot of it boils down to issues of system quality and loading as to whether
any of this matters (so, hey, on my workstations it does not matter because
the disk is not a big chokepoint; put 50 people running jobs on a cheap
cluster writing to disk via ntfs and believe me, you will notice).
Regards - Bob
----- Original Message -----
From: "Jason Swails" <jason.swails_at_gmail.com>
To: "AMBER Mailing List" <amber_at_ambermd.org>
Sent: Tuesday, December 15, 2009 12:18 PM
Subject: Re: [AMBER] Number of Cycles

> On Tue, Dec 15, 2009 at 12:13 PM, Carlos Simmerling
> <carlos.simmerling_at_gmail.com> wrote:
>> ah I see I thought Bob was talking about frequency of dumping coordinates
>> of
>> any type. I'm not sure why changing ntwr would affect performance but
>> ntwx
>> would not- is the the open/close on the restart file, or the
>> communications
>> to collect all coordinates?
>> thanks
>> carlos
>
> I believe he said writing mdcrds does affect performance, but so does
> ntwr. However, ntwx is needed for analyzing results, so playing with
> this is not a very good idea. However, writing a trajectory frame is
> cheaper than writing a restart file because a restart file uses higher
> precision and includes velocities (doubling the size of the file).
>
> Also, while the restart file is writing, I'm guessing the non-master
> threads are waiting at a barrier? In any case we don't have to
> sacrifice performance to write restarts, as they're not terribly
> useful, but we need to sacrifice performance to write mdcrds, as
> they're the source of our sampling, thus it's a necessary expenditure.
>
> Perhaps I'm missing something.
>
> All the best,
> Jason
>
>
> --
> ---------------------------------------
> Jason M. Swails
> Quantum Theory Project,
> University of Florida
> Ph.D. Graduate Student
> 352-392-4032
>
> _______________________________________________
> AMBER mailing list
> AMBER_at_ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber
>
>

_______________________________________________
AMBER mailing list
AMBER_at_ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber