RE: amber-developers: amber performance

From: Yong Duan <duan.ucdavis.edu>
Date: Wed, 28 Feb 2007 19:08:59 -0800

Hi, Bob,

Indeed, please focus on the important things. When you benchmark next time,
you may consider this ntwx thing.

yong

-----Original Message-----
From: owner-amber-developers.scripps.edu
[mailto:owner-amber-developers.scripps.edu] On Behalf Of Robert Duke
Sent: Wednesday, February 28, 2007 6:46 PM
To: amber-developers.scripps.edu
Subject: Re: amber-developers: amber performance


Okay, a couple of points on this writing stuff. For factor ix, a 1psec
trajectory interval is every 667 steps. Now, the pmemd 9 benchmarks for
factor ix and hfva (which I can't release in any case) are the ones that use

a 250 step write frequency; jac doesn't write anything, and if you look at
the benchmarks I added in the ORNL and Amber Dev talks, they are writing
with ntwx=1000, and this is a 2 fsec step. Now, if you want this stuff to
be fast, you use NetCDF on a machine that has a good file system like the
sp5/gpfs. Doing that you get a jump from 15.4 to 16.6 nsec/day on 320 sp5
processors for factor ix, even with ntwx=250 (so here is an 8% advantage to
exploit - but it may not work on something like an xt3 because the lustre
filesystem is a work in progress, putting it kindly). So back to data from
the talk. The master file i/o cost slide shows writing takes about 0.82
msec out of a 8.4 msec step. SO doing it 1/4 as much should save you about
7% for formatted i/o. These are not huge factors, but they help. I really
am not in a mood to back off and rerun all this stuff now to show this small

gain though. I think it much more important that I get some of the
functionality you all want yesterday, and then take care of the biggest
bottlenecks we currently face; if I don't keep moving, and somebody else
does, they will gain on us (geez, that does get old after a while...). So
the big problem with a unified set of benchmarks that everybody can run is
often just an issue of just being able to get the prmtop or it's equivalent.

So namd uses Apo1a, which has a bunch of residues that are a pain to do on
other systems (this benchmark is roughly equivalent to factor ix). I on the

other hand use a series of blood coagulation factors because Lee's lab
produces these things, so I can get them easily. But I can't expect others
to model these because they have some modified residues (gamma
carboxyglutamate among others). Also, past factor ix the various bcp's I
have are not yet available for general release because work on them is not
completed (ie., papers to come)). So the trick is to pick stuff that can be

easily used by several packages - ie., the standard amino acids.
Regards - Bob
----- Original Message -----
From: "Adrian Roitberg" <roitberg.qtp.ufl.edu>
To: <amber-developers.scripps.edu>
Sent: Wednesday, February 28, 2007 6:32 PM
Subject: Re: amber-developers: amber performance


> Robert Duke wrote:
>> Yong -
>> Don't fret about the ntwx 250; it is a matter of history, and as we move
>> forward we can change it to look more competitive. I have kept the
>> values so far in the factor ix benchmark for backward compatibility with
>> older results. Also, by cranking the write frequency up, I actually am
>> actively looking to push the disk i/o a bit harder to see where it will
>> be a bottleneck (so if I can get away with ntwx 250 at 300 processors,
>> then I should be able to get away with ntwx 1000 on 1200 processors). I
>> do feel that it is really only a fair benchmark if you DO write a
>> trajectory though, as most folks want at least a trajectory. If we
>> handle our file i/o efficiently, then we can end up having a competitive
>> advantage here (so the guys competing against us need to write
>> trajectories too).
>> Regards - Bob
>>
> I agree with Bob, BUT (big one also) the 'guys competing against us' might

> not be writing as often, and it is unlikely we can convince them to run
> fair benchmarks.
> So, it would be cool if you could rerun the benchmarks writing every 500
> or 1000 steps and tell us what they look like, so we can show off your
> stuff !
>
>
> --
> Dr. Adrian E. Roitberg
> Associate Professor
> Quantum Theory Project and Department of Chemistry
>
> University of Florida PHONE 352 392-6972
> P.O. Box 118435 FAX 352 392-8722
> Gainesville, FL 32611-8435 Email adrian.qtp.ufl.edu
>
============================================================================
>
> To announce that there must be no criticism of the president,
> or that we are to stand by the president right or wrong,
> is not only unpatriotic and servile, but is morally treasonable
> to the American public."
> -- Theodore Roosevelt
>
Received on Sun Mar 04 2007 - 06:07:29 PST
Custom Search